Thursday, November 19, 2026

The Majorana Door in Erice

The streets of Erice, Sicily are steep and narrow. Along one of them, a stone arch carries raised letters: Centro di Cultura Scientifica E. Majorana, Istituto S. Rocco. A heavy wooden door studded with iron sits below, and a tile marks the number 26. I stood there longer than a tourist should, because that name appears in my quantum writing.

Entrance to the Istituto S. Rocco, Ettore Majorana Centre for Scientific Culture, Erice, Sicily.

Erice sits on top of Mount Erice, about 750 meters above sea level, above Trapani and the western Sicilian coast. Antonino Zichichi founded the Ettore Majorana Centre in Geneva in 1962, and the charter carries the signatures of Bell, Blackett, Rabi, Weisskopf, and Zichichi. The Centre moved into the medieval town the following year. The Istituto S. Rocco is a former monastery, now the Isidor I. Rabi Institute. It houses the Directorate, the main secretariat, and the Richard P. Feynman Lecture Hall.

Zichichi was born in Trapani in 1929 into an ancient family from Erice. He saw the 750-meter hilltop as an ideal place for specialised meetings with a limited number of participants. Isolation, archaeological sites, and 3000 years of history added to the appeal. The town has cyclopean walls with Phoenician inscriptions and fragments of a celebrated temple. Zichichi began the yearly subnuclear physics meetings in 1963, and the Centre later took the name of Ettore Majorana, a young Sicilian physicist who had worked with Fermi in the 1930s. He also treated science as a tool for peace. In 1982 he promoted the Erice Statement, an appeal to scientists to serve peace rather than war. Zichichi died in 2026.

The Foundation reports that 112 people won the Nobel Prize after taking part in its activities, and 50 were already laureates when they first came to Erice. That adds up to 162. The same history page describes an IBM Fellow in Zurich who attended a school, stepped down from his management post, and pursued an idea that led to the discovery of high-temperature superconductivity.

Ettore Majorana was born in Catania in 1906. He wrote only nine papers. The last one, published in 1937, predicted fermions that are their own antiparticles. He proposed that the neutrino could be one of them. On 25 March 1938 he boarded a boat from Palermo to Naples and vanished. His fate remains uncertain.

The neutrino question remains open. The Majorana Demonstrator at Sanford Lab searched for neutrinoless double-beta decay, a rare decay that bears on whether neutrinos are their own antiparticles.

Quantum computing uses a related idea. Majorana zero modes appear at topological defects, such as vortices, boundaries, and domain walls in topological superconductors. In a nanowire design, they sit at the wire ends. Four modes form one qubit, and a set of 2N modes has a 2^(N-1)-fold degeneracy that local perturbations cannot lift. The information lives in the modes jointly, so local disturbances cannot act on it. Swapping, or braiding, the modes performs gates. Braiding supplies the Clifford gates, which form only part of a universal set, so designs add other operations.

Experiments followed the theory. Microsoft contacted the Delft group in spring 2010 about the Majorana hunt. In April 2012, Mourik and colleagues reported bound states pinned at zero bias in indium antimonide nanowires coupled to a superconductor. The authors wrote that the data support the Majorana hypothesis.

Microsoft announced Majorana 1 in February 2025, with eight topological qubits built from indium arsenide and aluminum. Physicists at the APS Global Summit met the claim with skepticism, and Science reported that many researchers saw no solid evidence of Majoranas yet. A peer-reviewed critique in Nature alleges coding errors and selective data presentation. Microsoft says the errors are trivial and the physics is sound.

On 2 June 2026, Microsoft introduced Majorana 2. The company reports parity lifetimes beyond 20 seconds in an indium arsenide and lead device, up from 1 to 12 milliseconds. Parity is one of two complementary measurements a working qubit needs, so the debate continues. In September, Microsoft gave DARPA on-site access to a system with Majorana 2 qubits at a new research center in Maryland. Independent testing has yet to establish how it performs. Microsoft targets a scalable system by 2029.

I stood under that arch with the steep street behind me. Majorana left nine papers and no explanation. Eighty-eight years later, his name labels a door in Sicily and a quantum chip. DARPA is currently testing the chip. The climb to that door is worth it.

Wednesday, November 18, 2026

Quantum Software Part 2: Testing and Debugging Quantum Software

This is Part 2 of a 3-part series on quantum software. You write an automated check, called a unit test, for a quantum program. The test runs a two-qubit circuit 1,000 times (called shots) and expects both qubits to read 0 in exactly 500 of those runs. It fails on almost every run, because the chance of landing on exactly 500 is about 2.5 percent. The program is working correctly and the test is written for the wrong kind of output.

Part 1 explained why the answers vary. This post covers how you test and debug software that behaves this way, and how it adapts to different machines.

A small example uses two qubits. The program links them so they always give matching answers when you measure them, either both 0 or both 1. The circuit takes four lines of Python in Qiskit, an open-source toolkit for writing quantum programs.

First, run it on a perfect simulator, which is software that mimics a quantum computer with no errors. Use 1,000 shots. The first run gave 491 results of both 0 and 509 of both 1. The second run gave 517 and 483. The counts change each run, and the pattern holds: about half and half, with no mismatches. A mismatch would mean one qubit reads 0 and the other reads 1.

Real machines make errors. The next numbers come from a software simulator, not a physical machine, set to assumed error rates of 1 to 3 percent on each operation. Across three runs, mismatches appeared in 6.9, 4.4, and 6.0 percent of shots. In the first run, 69 of 1,000 shots were mismatches. These random errors are called noise, and you can measure them.

Testing works on the pattern. A simple test checks that the qubits match in at least 90 percent of shots. The noisy run passed with 93.1 percent. A test that demands both qubits read 0 in exactly 500 runs fails on nearly every run.

Knowing the answer is right depends on the problem. Some answers are easy to check. If a program proposes two numbers that multiply to make a large number, you multiply them on any computer to confirm. Other circuits give the correct answer the highest odds, as Part 1 showed, so you run many shots and take the most common result.

More shots make that result more reliable. Quadrupling the shots roughly halves the statistical error, which is the random wobble in the counts. On a 50/50 result, 1,000 shots gives a margin of error of about 3 percentage points, and 4,000 shots gives about 1.5. You can also run the program on small versions of the problem that a classical computer can solve, then compare. Passing small cases builds trust and does not prove the large case. A fourth case, a problem with no known answer, comes up in Part 3.

Software can reduce errors. Techniques called error mitigation often run extra versions of your circuit and combine the results into a cleaner estimate. Each extra version requires its own shots, so better accuracy costs more runs. Chapter 12 of the book covers a related effort: AI tools that tune quantum hardware and fix its errors.

Debugging gets harder. In regular code, you add a print statement to see a value. Looking at a qubit means measuring it, which gives you a single value, 0 or 1, and ends that run of the calculation. The no-cloning theorem forbids making a perfect copy to look at instead. Researchers use indirect methods, and those add effort.

Machines also differ. Each one supports only certain operations, and only certain pairs of qubits can interact. A tool called a transpiler rewrites your circuit to fit. IBM's transpiler guide shows a small two-qubit circuit rewritten into a different set of operations for IBM hardware. You write a few operations, and the machine may run a longer sequence of simpler steps to get the same result. In October, I wrote about a faster version of the Qiskit compiler, which IBM reports transpiles two to three times faster on its benchmarks.

Return to the test that demanded both qubits read 0 in exactly 500 runs. A check that the qubits agree in at least 90 percent of shots passes on the same program, and it keeps passing when the counts change.

This is Part 2 of 3. Part 1 covered how quantum computers calculate. Part 3 will cover which problems justify the effort. 

My free book, Quantum from the Ground Up, is in this shared folder.

Tuesday, November 10, 2026

Quantum Software Part 1: How Quantum Computers Calculate

This is Part 1 of a 3-part series on quantum software. Part 1 covers how quantum computers calculate, Part 2 covers testing and debugging, and Part 3 covers which problems quantum software suits. 

Let’s say you write a simple computer program in Python that adds two numbers. You run the same two numbers 1,000 times. You get the same answer 1,000 times. Now you write a small quantum program and run it 1,000 times. The answers vary. The program is working correctly. Varying answers are how quantum computers report results. The program sets the odds for each possible answer, and a well-built program gives the correct answer the highest odds. You run it many times and keep the answer that shows up most often. The rest of this post shows how that works.

A regular computer stores information as bits. Each bit is a 0 or a 1, and the computer follows your instructions one step at a time. The same input gives the same output every time. A quantum computer stores information in qubits. A qubit can hold both possibilities, 0 and 1, until you measure it. The measurement forces one answer. Because of this, a quantum program sets the odds for each possible answer. The program is a list of operations called a circuit. Chapter 4 of my book, Quantum from the Ground Up, covers qubits in more depth.

Each run gives you one answer, so you run the circuit many times. Each run is called a shot. The pattern across all the shots is your result. If a circuit should give answer A 70 percent of the time, you need enough shots to see that 70 percent. One tutorial uses 8,192 shots to estimate a single value. On real machines, shots cost time and money, so you plan how many to use.

Here is how the calculation itself differs. A regular computer holds one pattern of bits at a time. Three bits hold one pattern, such as 101, and each operation turns it into one new pattern. Testing all 8 possible patterns takes 8 runs. Three qubits are described by 8 numbers called amplitudes, one for each pattern from 000 to 111. 

Each amplitude can be positive or negative (or more generally, a complex number with a phase), and one operation updates all 8 together. The count doubles with each added qubit. Ten qubits need 1,024 amplitudes. Fifty qubits need about 1.1 quadrillion, which takes roughly 9 petabytes to store on a classical computer.

When you measure, the chance of reading each pattern equals its amplitude squared, and you read one pattern per shot. Amplitudes that lead to the same pattern add together, and a positive amplitude cancels a negative one. This is called interference. A quantum algorithm arranges its operations so the amplitudes for wrong answers cancel and the amplitudes for right answers add up. The right answer then shows up in most shots.


The smallest example uses one qubit and one operation called a Hadamard gate. The qubit starts at 0. After one gate, the amplitudes for 0 and 1 are both 0.707, so a measurement gives 0 or 1 with 50 percent odds each. Apply the same gate a second time. The amplitude for 1 receives +0.5 along one path and -0.5 along the other, and they cancel to 0. The amplitude for 0 receives +0.5 and +0.5, which sum to 1. A classical random bit randomized twice stays at 50 percent, because probabilities only add. Amplitudes can cancel, so the qubit reads 0 every time. In a simulator with 1,000 shots, one gate gave 482 zeros and 518 ones. Two gates gave 1,000 zeros and 0 ones.

Varying answers still lead to one correct answer, because the odds are designed. Take Grover's search over 8 items, where one item is correct. Before the algorithm runs, each item has a 12.5 percent chance. After one round of its steps, the correct item has a 78 percent chance and each wrong item about 3 percent. In a 100-shot simulator run, the correct item appeared 76 times and the most common wrong item appeared 7 times, so the winner was clear. A second round raises the correct item to 94.5 percent. A third round drops it to 33 percent, so the algorithm needs the right number of rounds, which is 2 for 8 items. Noise lowers the correct item's share and raises the others. The correct answer still wins while it remains the most frequent result, and more shots widen the margin. Quadrupling the shots roughly halves the statistical error. For simulations of molecules and materials, the answer is the average across thousands of shots, and the spread becomes the error bar.


Let's go back to that Python program. Run it 1,000 times and you get one answer 1,000 times. Run the quantum program 1,000 times and you get a spread, and the program was built so the correct answer leads it. You read the pattern across shots, and the pattern is the result.

This is Part 1 of 3. Part 2 covers testing and debugging quantum software. Part 3 will cover which problems quantum computers help with. 

My free book updated quarterly, Quantum from the Ground Up, is in this shared folder.

Sunday, November 8, 2026

IBM, Oak Ridge, and AMD Put the Decoder Next to the Qubits

Chapter 12 of Quantum from the Ground Up covers NVIDIA's Ising Decoder, which ran 2.5 times faster than PyMatching with 3 times better logical error rates. I wrote about that NVIDIA AI system for drifting quantum hardware in June. Chapter 3 separates physics that works from engineering that makes it work. A November 5 announcement from IBM, Oak Ridge National Laboratory, and AMD adds a data point to both chapters. This post explains the announcement from the beginning, so you can follow it without a physics background.

Why Quantum Computers Need Error Correction

A qubit holds information in a fragile quantum state. Heat, stray electromagnetic fields, and small imperfections in the control signals all disturb that state. Each disturbance introduces an error. A useful calculation needs millions or billions of operations, so even a small error rate per operation ruins the result.

Quantum error correction addresses this. It stores the information of one logical qubit across many physical qubits. Extra qubits check the group repeatedly. Their measurements, called syndromes, show where an error probably occurred without revealing the stored data. Chapter 4 covers the basic idea. My post on two fault tolerance results in one week defines these terms in more detail, along with code distance and the color and surface code families.

What a Decoder Does

A decoder is a classical program. It reads the syndromes and decides which correction to apply. It runs on conventional hardware, not on the qubits. I described this classical computing layer inside every quantum computer in June. Qubits hold their state for only microseconds to milliseconds, so the decoder has a tight deadline. If it falls behind, errors pile up faster than it can fix them. That post reports that decoders run on FPGAs or ASICs for this reason. An FPGA is a chip whose circuits can be reconfigured for one job. An ASIC is a chip built permanently for one job.

Latency is the delay between sending data and getting an answer. Every cable and network hop adds some. A decoder in a distant data center adds latency to every round of measurements, which makes the deadline harder to meet. The same June post quotes Riverlane's argument that GPUs add latency even when their throughput is high. A GPU is a processor built to run many calculations at the same time. That argument sets up the November 5 announcement.

What IBM, Oak Ridge, and AMD Built

IBM, the Department of Energy's Quantum Science Center at Oak Ridge, and AMD unveiled a co-design platform for quantum error correction. Co-design means the quantum and classical hardware are developed and tested together. The first deployment sits at IBM's Poughkeepsie, New York site. It co-locates the 120-qubit Nighthawk with Umbra, an Oak Ridge supercomputer built on AMD Instinct MI355X GPUs and 5th Gen EPYC CPUs. Co-location means both systems share one facility. The Quantum Computing Report describes the design as aimed at eliminating long-distance wide-area network (WAN) latency between the quantum control hardware and remote supercomputers, relying instead on direct high-speed local interconnects. Researchers can test error correction on real quantum hardware without building that infrastructure themselves. IBM also joined two cryogenic modules in August and plans to install Nighthawk processors in 2026.

A decoder in a separate facility compared with the co-located platform. Simplified illustration.

Nighthawk gives researchers square-lattice connectivity and real-time pulse control. Square-lattice connectivity means each qubit links to its neighbors on a grid. That layout suits many error correction codes. Pulse control means researchers can adjust the signals that operate the qubits directly. IBM also names hierarchical codes as a promising approach. These codes stack several layers of correction, and they depend on strong classical compute. The platform lets researchers test those codes with GPUs in a low-latency setting. The announcements I found report no decoder timing results yet, so the question of whether GPUs can meet decoding deadlines at this scale stays open. One analyst firm frames error correction as a competition to supply classical silicon.

How It Fits With Other Results

Quantinuum reported 99.9995 percent logical fidelity on Helios on October 19. Logical fidelity measures how often a logical operation finishes without an error. At 99.9995 percent, about 5 operations in a million fail. IBM Research Director Jay Gambetta said the open task is speeding up and optimizing error correction. The Department of Energy's Quantum Genesis initiative targets fault-tolerant systems by 2028. My post on the $215 million Q Competition covers its requirement of at least 100 logical qubits. The partners say this platform supports that goal.

Terms to Know

•       Physical qubit: an actual hardware qubit, such as a superconducting circuit or a trapped ion.

•       Logical qubit: one error-protected qubit built from many physical qubits.

•       Syndrome: a measurement that shows where an error likely occurred.

•       Decoder: classical software and hardware that turns syndromes into corrections.

•       Latency: the delay between sending data and receiving the result.

•       Co-location: placing two systems in one facility to shorten the path between them.

What This Changes in the Book

Chapter 12 treats decoding mainly as a software and AI problem, with NVIDIA's Ising Decoder as the example. Add a hardware dimension. AMD hardware now supplies the classical compute for decoder research at Oak Ridge, and IonQ runs a real-time decoder on one standard CPU in simulation. The chapter should also note the distinction between microsecond-level real-time decoding—where FPGAs and ASICs remain the established choice—and near-real-time or multi-layer hierarchical decoding suited for GPU clusters. Chapter 3 gains a current example of engineering work, since the platform exists to answer integration questions. Chapter 6 gets a companion figure from Quantinuum. Logical fidelity and the 99.99 percent two-qubit gate fidelity cited for IonQ measure different things, so the chapter should list them separately, as I noted in October.

This post will appear in the next edition of Quantum from the Ground Up. You can download the current edition from the second edition announcement.


Thursday, November 5, 2026

Louis at Plainville Casting, Westfield, 1976 and 1977

I am slowly writing about people who influenced my life, so my family and the generations that follow can know them. I wrote some of it before my father passed away, and I have written more since. Louis was from Puerto Rico. We had little in common but we connected.

Pillow Block and Bearing
I started as a grinder at Plainville Casting in Westfield and then had various jobs in the factory depending on who was on vacation. As a grinder, castings came to me with extra iron on them and I ground it off. The parts ran from a couple of pounds to fifteen. I ground hundreds a shift. My hands buzzed when I clocked out and I was covered in iron dust.

I worked there the summers of 1976 and 1977 and one January in between. We cast anything made of iron that weighed fifteen pounds or less. Bearing housings and pillow blocks. Gearbox and pump housings. Pulleys, flywheels, handwheels. Valve bodies and pump impellers. Machine bases, brackets, covers, clamps, levers. A few decorative pieces. I still dream sometimes about the bins of Fafnir pillow blocks. In the dream I am grinding, the bins keep coming down a track and are always full.

The crew was Polish, Puerto Rican, and a few Russians. Every summer a handful of college kids showed up to save money for school. The language on the floor was Spanish and Polish with English where it was needed. I learned good words in all three. Most of them I cannot repeat here.

Louis was Puerto Rican and ran one of the sand mold machines. He spoke English with a strong Spanish accent. Five or six molders worked at a time. It was noisy, dirty work. The machines shook the floor. I could never believe how fast the molders worked. The cores came to them in a cart from the core room. Some arrived broken, and the broken ones went into a recycle can. I never saw any of them drop one. Louis pulled an overhead handle and sand dropped into the drag, the bottom half of the mold. He set the bottom pattern in the sand and the machine squeezed it tight. He lifted the pattern out and set a good core in the cavity by hand. He set the top pattern in the cope, the top half of the mold, and sand came down from an overhead hopper to fill it. The machine squeezed that too. He lifted the top pattern out and closed the mold. Then he passed the mold on to a pouring deck track. Then he did it again, and again all day long.

He called me Gordo. Everyone else called me Gordie. Gordo is Spanish for fat. I weighed maybe 135 pounds soaking wet. Louis was the first to say it, and it stuck.

He went to Wonderland Greyhound Park in Revere at least once a week. Sometimes he went with friends from outside the foundry. Sometimes he met his pals from Plainville Casting there. He read the program closely and he knew the dogs. One week he won. The next week he was broke. The week after that he won again.

A lunch truck pulled in every day. The crew lined up with their hands still gray from the iron. When Louis had won, he bought lunch from the truck. When he had lost, he brought a sandwich from home. You could read his week at the track in what he ate.

On the weeks he won, his wallet was fat with cash.

Other men played the dogs too. They talked about it at lunch. Who hit, who got burned, who owed who. Most of them told it as a joke. Louis never joked about it.

On nice days in the summer I sometimes sat with Louis on a bench next to the coke pile. The cupola took the foundry's own rejects, gates, and runners, along with old castings, radiators, and machine parts. Coke and limestone went in with them. The iron came out as castings. The pile was black.

We talked about sports, mostly the Red Sox, since it was summer. He talked about his wife and his children. I never met them.

The week he won $17,000, he showed me. He took out his wallet on the bench. It was packed with bills, folded over. He did not count them. He held it open so I could see.

“Look, Gordo,” he said.

I looked. I had never seen that much money in one place. I believe it was more than he made at the foundry in a year. He closed the wallet and put it in his back pocket. Then he went to the truck and bought lunch.

Two weeks later he was eating sandwiches from home again.

The molds came off the same either way.

Louis only talked about quitting after a bad stretch. On the weeks he won, he never mentioned it.

I have never gambled. I have bought maybe five or six lottery tickets in my life, all of them gifts. I never bought one for myself. One day he had a sandwich, and I asked him straight.

“You lose a lot,” I said.

“Sometimes.”

“Then you win.”

“Then I lose it again.” He turned the sandwich over and looked at it. “It is a problem. I know this.”

“Why go?”

“I try to stop. Four times. Five. One week I stop. Two weeks, one time. Then Friday comes.”

“What happens Friday?”

“I say I only go to watch. Then I am there and the program is in my hand.”

“Does anyone tell you to quit?”

“Everybody. My wife. Friends here. It does not help. I tell myself the same thing.”

He ate. A couple of the guys would toss a baseball around at lunch. We watched them.

“Gordo,” he said. “You go to school.”

“Yes.”

“Then listen. Do not gamble. Ever.”

“I do not gamble.”

“Good. Not one time.” He pointed the sandwich at me. “The first time you win is the worst. You think it is easy. You think it will happen again.”

“Did it?”

“Sometimes. That is the problem.”

I said I would not start. He nodded. He told me the same thing more than once that summer, sometimes in English and sometimes in Spanish. No apuestes nunca, Gordo.

“And you stay in school, Gordo. No matter what.”

The whistle blew and we walked back in.

The foundry is gone. I do not know what happened to Louis.

I never started. I stayed in school, as he told me to. I earned my degrees, worked for a while, and I became an educator. I have told students to stay in school ever since.

Louis knew what the track cost him. He said so to a college kid he had known for a few weeks. Nobody made him. I learned from him that a person can understand a problem completely and still not stop, and that they can still warn someone else.

Years later I put the name he gave me on my website: gordostuff.com

I still think about him.

Wednesday, November 4, 2026

What the 2026 Midterms Mean for AI, Data Centers, and Quantum

I have been writing about AI, data centers, and quantum for the past year or so. Here's a quick what I see ahead for all three.

AI

AI policy has a bipartisan draft on the table. Reps. Obernolte and Trahan released the Great American AI Act discussion draft in June. It would freeze state laws on how AI systems are built for three years. Developers with over $500 million in revenue would publish risk frameworks, submit to audits every six months, and face penalties up to $1 million per violation per day. AP counts on Wednesday morning showed Democrats ahead in both chambers, which makes preemption without those federal rules unlikely. Any bill still needs 60 Senate votes and a presidential signature.

Data Centers

Data center costs reached campaigns in both parties, as S&P Global reported before Election Day. The Atlantic Council notes that data center grid demand could roughly triple by 2030 and that utilities filed $9.4 billion in rate increase requests in early 2026. A bipartisan Senate permitting bill would cap legal challenges to federal permits at 150 days and require data centers to help pay for transmission. The House Republican Ratepayer Protection Act is voluntary. Expect Democratic committees to push mandatory cost rules and hold hearings on the companies building these facilities.

Quantum

Quantum has the least partisan path. The House Science Committee approved H.R. 8462 on April 29, and Senate Commerce passed its version unanimously. The House bill adds NASA as a partner and expands workforce and supply chain programs. I found no floor vote yet. The lame-duck session could pass it before January 3, or it moves to the new Congress. Funding depends on the budget fight, and Rep. Lofgren has already flagged cuts to science agencies.

The lame-duck session before January 3 gives us the first signal on all three. Watch which bills reach the floor. I’ll be watching and writing more here.

Monday, November 2, 2026

Second Shift Summer 1978

Donnie got me the job. He was already on the bandsaw when I came in that May of 1978. We were both twenty-one. I had spent two summers at a cast iron foundry, where the work was hot and dirty and a kid could get hurt. Savage was an improvement and it paid more.

I could walk to work. I cut through the woods behind my parents’ house to the railroad tracks. Right led to my trout stream. Left led to the plant. I wrote about that summer here when I turned left.

On my first day I had to buy a pair of steel-toed boots. The plant required them. We also had to wear safety glasses with side shields. The glasses were free.

The first week they put me on days with Smitty, who trained me.

Smitty was in his early forties with forearms like fence posts. He had come to the plant right out of high school and had never worked anywhere else. He did not talk much. He took a stock from the cart and held it up so I could see the grain.

“Belt first,” he said. “Then the block. Then the fine paper. Use your thumb.”

He ran his thumb along the wrist and the comb and handed the stock to me. I did it. My thumb caught on a ridge by the comb and I sanded it down and he nodded.

They paid for what you finished. Nobody told me how many I should do. I counted my own. The first day I did thirty. Smitty did eighty and never seemed to hurry. On Friday he watched me work for a minute.

“You hold it too tight,” he said. “Let it go where it wants.”

Then he went back to his bench.

The next Monday I moved to second shift. Smitty stayed on days, so I did not see him after that, except once in a while in the lot when he was leaving and I was coming in. He always raised a single finger off the steering wheel with a big smile on his face. I always reciprocated with my own single finger reply

The shift started at three. By three-fifteen the shop was loud and the air was brown with sawdust. I had a section near the back. I took a stock from the cart and put it to the belt. Then the block. Then the fine paper.

By July I did sixty a night and averaged twelve dollars an hour. The minimum wage was two sixty-five. At twelve an hour, a night came to ninety-six dollars and a week to close to five hundred. I had never held that much money. I did the arithmetic on the walk home and thought about what it would buy. Dust sat in the creases of my clothes, in the seams of my steel-toed boots, under my fingernails, in my hair.

In July Smitty took two weeks of vacation and they put me on first shift. I walked to work in the morning and home in the afternoon, in daylight for once. On one of those afternoons I came home with sawdust in my clothes and told my mother I was quitting college to sand gunstocks full time. She said no.

Donnie’s work was louder than mine. He fed blanks to the bandsaw all night and the blade ran with a high steady whine. One Tuesday in July, at break, he came over with sawdust on his shoulders and his ears ringing, and he shouted until his hearing came back.

“Come on,” he said.

We went out the back door past the loading dock. The air was cool after the shop. Behind the building stood the water tower, and past the tower ran the Massachusetts Turnpike. We heard the trucks and cars before we saw them.

The ladder started high enough that Donnie had to jump for the first rung. I went second. The rungs were cold and the paint came off on my palms. We climbed without talking. If anyone had looked up from the dock we would have been fired. Nobody looked up.

At the top a narrow catwalk ran around the tank. We sat with our legs over the edge. Below us the plant was a long low building with yellow light in the windows. You could hear the machines through the walls, but quieter from there. The Turnpike ran east and west, and the trucks went by with their running lights on, one after another, going to Albany or going to Boston.

Donnie lit a cigarette and cupped it in his hand so the light did not show. I never smoked.

“Sixty,” he said.

“Sixty.”

“Smitty does eighty.”

“Smitty’s been doing it since high school.”

Donnie thought about that. He looked at the road.

“Twenty-four years,” he said.

Neither of us said anything after that. We sat until the bell.

Going down was worse. Going up I had watched the next rung and the next. Going down I could not see my feet. The tank blocked the light from the plant and the ground was somewhere below in the dark. I felt for each rung with my steel-toed boot and held the rail until I was sure of it. Donnie went first. I could hear his steel-toed boots on the iron and then I could not tell how far below he was. I did not look down. I counted rungs.

At the bottom my hands shook. I wiped them on my pants. Donnie said nothing about it. We walked back across the gravel with the dust still in our clothes. Inside, the belts were already running. I took a stock off the cart and put it to the sander. I held it loose.

After the shift I walked home on the tracks in my steel-toed boots. The rails ran straight in the dark. The tower stood against the sky behind me until the trees closed over it.

We climbed about once a week that summer, on the Tuesdays and Wednesdays when the nights felt long. Going down never got easier. Nobody ever caught us.

The tower was demolished over 20 years ago. I'm trying to find a picture of it. I'll post one here if I do. If you ever drove the Massachusetts Turnpike back then I'm thinking you remember it.

Donnie earned his degree at Springfield Tech and became an electrical engineer. He rode a Harley. He cooked with his daughter and watched his grandchildren on the weekends. Years later I ran a center at the school he attended. He died on a Friday in January of 2017. It was his heart. He was fifty-nine. I never told anyone about the tower. Don't think Donnie did either.