A Day in the Life of Tony

At Kensho, we pride ourselves in having a community of employees who come from all different walks of life and who fill a wide array of intellectually challenging roles.

Kensho’s Day in the Life, a series started by the Diversity and Inclusion committee at Kensho, gives readers the opportunity to glimpse into the day-to-day experiences of these employees and their journeys in arriving where they are today.

Tuesday, October 3, 2017

Don’t grind your teeth.

Tony opens his eyes to the eggshell-white ceiling. Gray dawn light streams through the blinds on the one high basement window at the foot of the bed, to the east. His alarm hasn’t gone off yet. He remembers no dreams, only a black lacuna. The house is quiet; the three others won’t be up for a while.

He thinks it’s genetic. The teeth thing.

He pulls his phone out from under his pillow and checks his email — newsletters from the New York Times, Wall Street Journal, Quartz. The news is grim.

Tony leaves his phone on the bed and walks upstairs to the kitchen and over to the AeroPress to brew. Without a watch or phone or clock he has to count two minutes fifteen seconds in his head. As he waits, he tries to lay out his day, imagining what he wants to get done and what promises he’s made. He’s got some food he wants to drop off at his family’s house. It’s been too long since he played music. He’s going to Nashville on Thursday night. So he waits and thinks and lets the coffee brew too long and takes a sip but it’s almost too bitter to drink.

Outside, ​the sidewalks are still empty but the bus stops are already full. The street sweeper loiters in the bike lane. A couple runs together past the gas station. Those who are walking at this hour are walking either with supreme purpose or none at all. It’s the time when Dunkin’ Donuts has long since opened and nothing else has.

A voice recorder shows up at the door. Tony lets it in and it sticks in his face. He goes downstairs to change, taking the stairs at a syncopated trot, tip-tap, tip-tap.

Normally he’d have been up earlier — he’s been trying for 5:30 — but he’d gone to sleep late last night, around 11. He’d been arguing with a Ph.D. statistics P.I. he used to work with about parameter estimation for autoregressive models.

When he steps outside it’s 7 a.m., 45 degrees — an early October local nadir, chillier than both the previous week and the next. The chickadee whistles FEE-bee above the long gravel path snaking around the house and out to the road.

Tony runs in the morning, three times a week. He has for a couple years. He’s energized by the routine. He never ran athletically, only taking up this morning curriculum after graduating in 2014. He runs down toward Harvard Square, looping around to the west just short of it. He runs past Science Center and then Maxwell Dworkin, where combined he spent the bulk of his undergraduate class time, to the northern edge of campus, which he never used to cross at all, and beyond, past Lesley.

In the home stretch, as he turns right onto Somerville Ave., he gets a view of distant downtown Boston, the Federal Reserve Bank washboard straddling the double yellow line. The skyline relieves some of the claustrophobia of having spent so long in the same part of Cambridge.

“I live so close to Harvard that otherwise I’d feel like I hadn’t gone anywhere. When I first got out of college I actually worked down there, so I would see the skyline as I was biking toward downtown.”

He worked for a little consulting company that got bought out by a big consulting company. He clustered customer data for retailers, or culled purchasing patterns from inventories, or tweaked broadcast schedules to improve ad revenue. They did good work, but didn’t have the clout for their recommendations to be totally compelling for clients, who sometimes abandoned projects for political reasons and left Tony unsatisfied. They used Java and did a lot of one-off unoptimized analysis and had only just discovered version control, and Tony yearned for more. And didn’t want the “more” to be ad optimization. And he was already living with three friends who worked for this fintech startup, and they recruited him.

Back at home, the house is still quiet. Tony takes off his shoes, blows his nose, and puts his shoes back on. He walks across the street to Starbucks and orders a pastry and a coffee and sits. He looks at the voice recorder.

“You see a lot of athletes described as people who are curt, or rude, or whatever. Well, no, they’re just, as compared to other people who are on TV all the time, they view their job as playing sports, right?” And the journalist has to be “comfortable with communicating to someone with your behavior that you’re here for them as a concept, and not as a person.”

Tony listens to sports podcasts to empty his mind, or to connect with society. He admires the transverse application of power, from athletic prowess to social advocacy, in the face of its risks, of how easy it is to be misconstrued. He likes to listen in the early morning. The morning, before his inbox wakes up, is time for listening, and for reading, and for quiet contemplation. The quietude recalls college.

“To grip an idea requires you to just stare at it for a very long time. One of the reasons I decided I would study math is because there’s this great math competition you can do in high school called the USA Math Talent Search, which is different from most others because it’s a time limit in the weakest sense: you have five problems in a month. There was one problem regarding a tetrahedron; I think you had to figure out something like the length of some line segment that was created by, you know, you take the perpendicular here and the perpendicular here, how far is that from the base? Or something like that. And in order to do it, I made myself a tetrahedron, and before I went to sleep I would just rotate it in my hand. And one night I solved it, half-asleep, just because I’d been thinking about it for so long, that I finally realized, literally: Oh, if I look at it from this angle, I can think of that right triangle as kind of like this right triangle, and those are actually similar, so I can do this thing and it all works out. But very literally, just because I was staring at it, thinking about it all the time.”

Tony walks out of the Starbucks. There’s a crosswalk 20 feet away, but Tony just steps off the curb and crosses the empty street at a 30º angle, equal to his angle of incidence, jaywalking with the confidence of a ray of light passing into a medium with an identical index of refraction.

Dew is evaporating off his mailbox. Tony showers and dresses and walks a block to the Hubway bikeshare station. There’s traffic on Oxford, so he takes Irving and Kirkland.

Tony gets in at 9:30. He bumps his butt against the keyfob sensor in the elevator and presses ③. He goes to his convertible standing desk and stands leaning against a sort of stool thing, a Varichair. It’s supposed to be good for posture. His desk has two large monitors: one connected to his MacBook, one connected to his siloed ML box. A small button on a dongle switches his keyboard and mouse input between machines. There’s headphones, a couple water bottles, a stack of ML papers, a stuffed animal. The keyboard is a Goldtouch V2 Adjustable, arched with an aggressive ergonomic split. He logs on, checks Twitter, reads a blog post, “TFX: A TensorFlow-based production scale machine learning platform”, and posts it in his team’s Slack channel.

His first task of the day is to finish up a data quality check on some new stuff his team got yesterday. The data’s in a SQL database he’s read into a pandas dataframe in Jupyter. He runs through some standard checks: what are the unique values, how many null, what’s the distribution, maximum, minimum. How does it look aggregated different ways. He checks if anything looks oddly large, or missing, or duplicated. He makes sure relationships between columns make sense. Currency units line up. Date and time columns agree with each other, with time zone expectations, with London and New York market hours.

Every fifteen minutes, an app locks his screen for one minute of eye rest.

Some errors are obvious and easy. A giant number in the data might wow him—like, finance is crazy!—until he realizes it’s a thousand times too big, a fat-finger duplicate of another row. In that case he can confirm with the data team on the other side. But some are less tractable, gaps and vagaries and biases.

“It’s a lot of guesswork and intuition about what ought to be consistent. I would say the errors tend to be subtle. A lot of the large things are correct, so you might not notice anything off, but you might question certain things like, why did there look to be slightly less trades in this year than other years? The answer might just be, there were less trades in this year than in other years. But the answer might also be because we’re missing data for this particular class of assets or from this particular region of the world, in that period, because of some data error that happened on their side. And it’s a rabbit hole, because you can have millions of hypotheses about why the data is right or wrong, and you have to stop investigating it at some point. And not all of them can be fixed. So it’s an evaluation of, to what degree does this need to be right before we can do something about it? Because fixing it might take six months.”

Tony likes the challenge of imperfect information, the sense of: Well, shit, we still have to model this correctly.

At 11:00 he interviews a job candidate, a current college senior. He asks things like, “Explain your favorite ML algorithm like I’m a college freshman.” It makes him miss school a bit, “seeing the opportunities available to the kids. The average level of material they’re exposed to and are aware of will always increase.”

He writes up his interview feedback, and then he writes a data quality report with his findings from the morning — observations about what data’s worth keeping, an appropriate query so his findings are reproducible, some exploratory code for the repo.

At midday Tony walks out the door to the stairwell. Now he takes the stairs at an even walk, left, right. He clasps his hands, stretches them out in front of him. He tosses his head back and laughs. Emerging onto Dunster St., Tony shouts over the sound of jackhammers and workers swarming the fenced-off Smith Campus Center.

“You know, I’ve never thought about the amazing sparsity of dialogue in some novels. When you read an old Russian novel, right, there’s a strong tendency toward lots of inner reflection. In, I forget which one, I want to say Anna Karenina, there’s a long time where this one character is just in the field mowing grass.”

Tony walks into Pinocchio’s and orders a small fried chicken sub and a Sicilian slice with spinach. No drink. $10.94. He walks back to the office and finds an empty room in which to eat.

Luis peeks in. “May I join?”

Luis comes in and sits down with his lunch. They chat; the conversation turns from shadowing coworkers, to the awkwardness of imposing, to grappling with focusing on the narrow scope of a story.

“I read a lot of sports news and follow a lot of sports journalism, and it’s their job to talk about sports and generate content so people can hear about sports. But yesterday, a lot of them woke up in the morning and were like, ah, I’m supposed to podcast about yesterday’s football game, but a lot of people just died, and I’m just going to be talking about… football? I guess I do this and despite the fact that I have a public platform, I ought not to spend too much time talking about it, because I’m here to do my job. But how do I acknowledge it, still, because it feels wrong as a human being to… not?”

The conversation turns to meeting schedules, to open office plans, to television shows. Tony rests his left foot on the rim of the base of his wheeled office chair and crosses his right over it. From his right hand he sips a Polar seltzer, unflavored; his left hand rests with his thumb to his forefinger.

“Did you watch the season finale?” Luis asks. Tony nods. “Did you like it? I thought it was really good, but the ending there…”

“There wasn’t enough suspense for you?”

“No no, it’s like: everything is back to normal but… is it? It’s not possible.”

Tony runs his fingers along the left armrest, then lets his arm fall to his side. “I’m trying to remember exactly how it…”

“There’s a lot of… in the last act — at least, in my interpretation — he keeps discussing, ah, maybe this is not real, I can just change worlds, nothing matters, I can go to another dimension, things are easy, I may be a simulation, I may be I-don’t-know-what — but then, in the end, everything has to go back to how it was in the first season?”

“I’ve always thought of it as kind of the — I guess what now must be like — I don’t know, it feels like a very basic philosophical notion — basic in the cultural sense, not in the literal sense of complexity — but maybe like the notion of Camus’s response to the nihilism of reality, which is just, you must imagine…”

“But we’re not sure because she may be a clone, and the real…”

“It’s true, but for all we know…”

“But I see your point…”

“But um…”

“It is the kind of thing where you are not sure if it is super smart or it’s just dumb.”

Tony smiles. “I mean, they don’t care.” Luis laughs.

Tony pulls his phone from his right pocket. He sees that they pushed a Thursday interview an hour later than it was supposed to be, which will make it harder for him to catch his flight to Nashville. He tosses his trash and goes back to his desk.

He reviews some code and fixes some bugs. “Naming conventions and formatting and all that jazz; it’s a very minutiae kind of day. I think one of the things I try to be sensitive of is to be a team player.” But, “there’s this tension between, every moment I spend making these aesthetic changes, I’m not working on the data or the model.”

He refactors some model generation code to be less duplicative. He reads some tests associated with it. He fiddles with a blue coiled hair tie he’d found on his desk one day, repeatedly folding it in on itself and aligning the notches.

He listens to music as he works. TOKiMONSTA, Giraffage, Anderson .Paak, vbnd, Covet, Mount Kimbie, Kelela, Moses Sumney, King Krule, Demi Lovato, Airhead, Philip Glass. He’s going to an Aquilo concert tonight at the Sinclair, $16, on the recommendation of a friend. It’s some kind of indie-pop-electronic band. He listens to some of their stuff on Spotify.

It gets hot in the office; Tony moves to the back room. “Saunters”, maybe. A friend from high school once described his walk as a saunter. Tony saunters with his laptop to the back room. Later Tony saunters over to Shleifer to talk about a big refactor to accommodate more data and scope. Tony saunters over to Harrison re: same.

Tony saunters to the kitchen and picks up some potato chips and dark chocolate. The voice recorder is there, asking Martin about dreams.

Martin says, “Imagine if instead of lucid dreaming we could do lucid bitcoin mining?”

The voice recorder asks Martin how it feels to live with Tony.

Martin says, “He’s just a beautiful person to be around. We have great chemistry. His intangibles are magnificent. We really think that he should be the starting quarterback of our team.”

Tony saunters away.

All code Tony writes goes into the version control system git and, atop it, the code review tools of Phabricator. In the musical composition of his life, Phab is a bass ostinato—repetitive, regular, fading into the background. Tony will write some code, probably Python, probably in a Jupyter notebook; at this moment it’s converting a string (i.e., unstructured text) to a list of lists (i.e., structured data) with ordered = [x[:-1].split('<') for x in s.split('; ')]. The code lives in a git repository in a folder on his computer, branching off from the most recent shared version like an alternate history of what the code could be. Then, at some natural stopping point, or when he needs feedback, or when a teammate needs that code, he “commits” that code to the repository — in this case, with git commit -am 'wip', where “wip” is a (somewhat lazy) comment by Tony that this is a work in progress. Committing gives a short name to this exact version of the folder, and saves it. The contents determine the name; it’d be identical for an identical folder, but totally different for even the most similar folder, so it shows if things have changed. It looks like: e83c5163316f89bfbde7d9ab23ca2e25604af290. And the commit gives credit to Tony and references the preceding commit. So then he “diffs” it using Phab’s command line tool Arcanist (arc diff), which pushes the code to a server and notifies his team. Designated reviewers (e.g. teammates) can then see that diff, i.e., see what Tony added, deleted, or changed, compared to the last shared version. There they can comment on individual lines, make suggestions, request changes, etc. Tony responds by adding commits to the diff with any new changes as needed. Then, if Tony’s code looks good, a reviewer accepts the diff, and Tony lands it (arc land). At that point, his commits are squashed into a single commit accumulating all his changes; the repository is rewound to the moment he branched off; then it’s fast-forwarded with his changes applied, as well as anything anyone else has done since then; and then the meaning of the repo’s master reference is updated to point to the name of the latest commit, so that if anyone references master, they’re all talking about the same version of the code, all up to date.

Depending on one’s familiarity with this common-but-specialized workflow, this may sound either arcanely incomprehensible or obviously banal to the point of not-even-worth-saying. The human impact is that Tony is bound up in (and sometimes blocked by) threads of reciprocal accountability, formalized in code: he needs his team to see what he’s doing and reach some incremental consensus on it, and they routinely need him to do the same.

Except, if Tony’s working on a solo project, as he is now, he can land without review. His team trusts him.

His last diff of the day is changing a column name from “security” to “sector”.

The moon is full as hell (95.88%) and low in the sky (9º 46' 40.7"), level and in line with the bulbs of the streetlamps and about the same size (subtending 31.306'), as Tony walks toward Clover (109º 41' 43", ESE). A woman walking the other way on Mass Ave. waves.

“Tony!”

“Oh hey!”

“How’s it going!”

“Not bad!”

“It’s been forever, good to see you! How are things?”

“I’m doing well!”

“That’s great.”

“How ’bout yourself?”

“Uh, doin’ alright, things have been”—the woman pauses—“I’ve had a really hilarious last few months, ’cause I, uh, I landed a job, at what I will admit is sort of a dysfunctional company but I knew that going in, and then everyone above me on the technology team quit!”

“Exciting!”

“In a terrifying sort of way, yeah! How ’bout you? What’re you up to?”

“I work at a tech company right in the square right over there.”

“Sweet, doing what?”

“Software engineering, data science, whatever.”

“I’ve gotta go meet my friend actually in Davis Square but it was good to see you, we should get caught up!”

“For sure!”

“See you around!”

At Clover, Tony orders a genetically modified meatless meatball sandwich and corn fritters. He dispenses a compostable cup of water and sits in a booth while he waits. When his sandwich comes, it tastes almost like meat.

He’s in good shape for tomorrow. He’s in control of his responsibilities to his team, he’s getting stuff done, hitting short-term objectives. Small fixes, making sure people aren’t blocked, cleaning up code. But he’s beginning to miss the data. He wasn’t working on the model, learning more about the domain, drafting mocks of new directions. It was all incremental changes. But he’s at peace with that. Sometimes he needs those days. It keeps him coming back, for the other days — the days when he gets to think more deeply.

“Finance is hard to model and a lot of things don’t work. But you can see both the internal logic and also the parts where the internal logic fails. It’s tricky when we want to model transactional behavior — someone might decide not to make a transaction today for no better reason than they were like, ‘Ehhh, I kind of have a feeling.’ Right? And when you’re affected by human whims to that degree… it’s a chaotic low-level rule-based system, instead of one governed by high-level rules. And on the flip side, there are people who are actively trying to deceive you. So, you know, ‘Are they stupid or are they smart?’ — I feel like is a question of our times.”

“I’m a lot less cynical about it now than I was before. I think actually, having studied a lot of its more complicated products in depth, I understand more their purpose, what they’re trying to do, how they’re trying to function. Obviously there’s a limit to that. By the time you’re making the nth-level derivative of some financial product, how much value you’re providing is truly questionable. But the first couple levels are pretty natural, right? Like, we want to make sure that we can get paid if our bonds default — that’s why you have a credit default swap. But does it make sense that people trade options on the VIX?” (Lauren later says: as a hedge to your equity exposure, it can.) “Well, maybe, maybe not; that seems like pure speculation. So I think my opinion of finance has improved a lot, because I can now do a better job of separating the notion of finance and speculation.”

After Clover, he goes back to the office to pick up his bag. It has: tickets to the concert tonight, a problem set on measures of bond risk, a couple notebooks, kleenex, glasses, and a drum machine. He has a little time before doors open, so he goes to Crema for a Moroccan mint tea and macarons, and works on the problem set.

This has not been a normal day for Tony. This morning he’d have run earlier and longer, if not for me. If I hadn’t been there, he wouldn’t have stopped at Starbucks. He’d have gotten to work at 8. He’d have eaten lunch alone, and quickly. He’d have gotten more done today, if only I hadn’t been seated next to him by chance on my first day, and Lauren hadn’t put out a call for people to write these, and I hadn’t thought first of him as someone I felt comfortable shadowing. And if he hadn’t lived with three people who already worked at Kensho — if he hadn’t gotten that interview with Analytics Operations Engineering, the October of senior year, and sat down to dinner with Martin, a classmate he hardly knew, and decided to live together in Somerville after graduation? If he hadn’t administered so many musical groups, if he’d studied more, if he’d gone to a less careerist school, gone on to grad school for math? If he hadn’t been such a solitary nerd in high school; gotten near-perfect standardized test scores; earned glowing recommendations from his English, Chemistry, and Fine Arts teachers? If he hadn’t found Asimov’s The Atom as a young child, mining the Dewey Decimal System for all the science books? If he hadn’t been born to a geneticist and a stem cell researcher who left bio textbooks lying around, the earliest books Tony remembers seeing, seeding his earliest imagination with macrophages and t-cells and antigens — born, in the United States of America, on a warm wet sunless summer day A.D. 1992…

There is no unbiased estimator of a day or a life. We cannot escape our sampling methods.

Tony meets his friend at the Sinclair when the doors open at 8, with an hour to kill. He gets a beer, rests his left elbow against the wooden moulding around a concrete pillar, chats about her work at the GSD. Sometimes Tony thinks about venues in NYC and wonders if he’s happy in Boston, so he makes an effort to make the most of what’s here, to keep him sane, even though it strains his strict routine. The opener is a painfully hipster banjo and drums from Brooklyn, playing songs about their pain and love and isolation and feelings. He idly twirls his yellow wristband. He thinks the banjo is superficial signaling, the wrong choice for the music they’re trying to play, too quiet. Like, get a guitar. Holding his drink in his left hand, he claps with his right against his wrist. Between sets he idly wonders why Gordon Gekko slicked-back hair was ever a thing. As Aquilo comes on, he’s debating whether he should care what Pitchfork says. By midnight he’s thinking he should tell his college roommate he’ll forever be his brother.

Tony goes home.

He lies in bed, thinks back to yesterday, to Las Vegas, thinks I imagine gun control is complicated, but I can’t help feeling like it shouldn’t be this hard. And at 12:45 a.m., like everyone does at some point or another, he closes his eyes. You already know what it’s like. His senses decouple from natural phenomena; the chain slips off the sprocket; images decohere into fancies and visions and voids, as his freewheeling consciousness dissolves in the froth. There is no world, there is no time, there is only Tony.

Tony Liu is a machine learning engineer on our Machine Learning team. Toph Tucker is a designer or developer or something on our KMI front-end team. Thanks to Lauren, Lelaina, Adity, A.B., Shleifer, Neary, Predrag, and Jessica for their prompting, feedback, and patience.