In a year, I received and answered roughly two hundred emails from people of divergent walks of life, sharing a seemingly universal pain: finding it hard to “make it” in ML research.
Their definitions of “making it” can be vastly different, their paths to research and positions in life can be all over the map, and the authentic need behind such a desire can be anywhere from pragmatic to idealistic: some want to learn something new and exciting, some want a real breakthrough after years of grinding; some want to make connections in the seemingly shiny AI circle, some find the circle they finally make it in foreign and alienating; some want publications in prestigious venues to gain recognition, renew or maintain reputation, graduate or be promoted, while some want publications of any sort to learn the game; some want a strong validation to keep pursuing their chosen path, and yet some, only a vague sign whether a path is worth choosing at all.
Some want to move up, some to move in, but all in all, they want to move forward and find it hard, for reasons we all know so well by now: the societally misconstrued promises, overly reductive credit attribution, archaic mentorship structure in academia, as well as the extreme competitiveness of a game (scientific publishing and hiring) never designed for a scale seen today.
The consequence is almost everyone, junior or senior, accomplished or not, lives with some measure of anxiety, fatigue, demoralization, and loss of meaning.
I know it because I’ve been through it myself, and I’ve heard it countless times, thanks to all those who trusted me with their stories. In the midst of 2020 when life took me to the darkest place I’ve ever been, I went blindly looking for resources online for help (but at the same time disbelieving anything could really help me), a sentence struck me: “In order to feel better you have to want to feel better.”
Yes, self-help, simple as that. In the same sense MLC is never meant to be your magic ladder to success, or the prescribed treatment to your anxiety; it doesn’t even claim to be your helper. It is simply a place for you to go, when you are ready to help yourself.
If you are at a point where “self-help” is not really the focus of your life, and you are just genuinely curious, you can read all the grandiose (but authentic) mission statements of MLC on our website, especially the About page.
But if you are at a point of wanting and needing to feel better, caring about, even to the point of worried about, your own wellbeing, like how I was when it seemed like every bad fortune was falling upon me and some demonic shadow was clouding my body, maybe this post about why I chose to build MLC the past year—from purely a self-interested point of view—would be of interest to you.
This is a personal retrospective of my year running MLC. It is “selfish” because I intend to only speak from a lens of self interest and self gain, as someone who had struggled to keep afloat, but was magically saved by pivoting to do something different, to stop chasing and start creating.
These personal takes would be most relevant if you are in ML research, and if you are struggling, but I hope they’d also make some sense if you are neither. For context, I start with a brief history of how I got here. You can skip ahead to Section 3 if you know me pretty well.
How it started
In 2016 I joined a tiny startup which fortunately turned into Uber AI Labs a year later. I (and most people from the startup) stayed long enough to witness the entire course of the rise and fall of such a venture. At AI Labs I worked with Jason Yosinski on likely a thousand things, but it is one thing he said to me in early 2017, when we were given the opportunity to start a team, that presaged the course of what MLC is today.
When discussing what this team should be like, he said, “I don’t want a team to be about Jason.”
This at the same time shocked and enlightened me. It was the first time I ever heard anyone say that; I mean, who wouldn’t want a thing—a team, a lab, a company—to be about themselves? How many labs, teams and companies out there are identified with, if not directly named after, a single most notable person?
But that’s Jason. He has always looked much further, and believed in a much more wholesome picture of what a research group should be. Even though he can (his reputation as a research lead even back then would’ve easily made us one of the most attractive teams in ML research), he is unwilling to take the egotistic path that the current society rewards the most.
That decision led to two results:
- We were never a traditional team. As a team of two (occasionally three or four counting interns and residents) our weekly team meeting could always easily fill a 12-person conference room, as we were open and inviting, and ended up attracting the smartest and most aspirational folks from all parts of Uber: research (in and out of AI Labs), data science, engineering, even policy and marketing. They became our teammates, even though they didn’t report to us. And they were passionate and committed enough to attend meetings, give presentations, lead projects, and eventually write papers. In the course of four years, all of our projects (which we now display under MLC: Projects), consist of coauthors, even first authors, who were not in our reporting chain at all. While researchers working together across teams might not be unusual in large research organizations, we are not a large organization and these are not team collaborations, but rather people from anywhere in the company coming on their own accord to do research with us. And a lot of them were writing papers for their very first time.
- I implicitly take up more responsibilities and shift my goals. Because it’s not a “team about Jason,” I am not just “someone working for Jason.” I am Rosanne, and that could mean whatever I want it to mean, and prompted my objectives to shift distinctly. I was allowed to take up leadership and management much sooner (without waiting for some title change or promotion to make it part of my job description), and I experimented motivating people without directly managing them, and found it useful to stay a generalist rather than burrow into some narrow specialty, in order to support people and projects of divergent interests. And none of this would’ve happened if we were a traditional team. Among these changes, the shift of objectives is the most profound. For example, as someone “working for Person A” one might want to first and foremost optimize her individual performance to get noticed by A. In the research world, that might translate to “publishing as many first-author papers as possible and branding yourself as an expert of a small research area” before you are promoted to oversee a team or org. But as “someone working for Team A” one’s goal can be easily purposed around optimizing the team output, so the team as a whole looks good. For me that was “making our team publish as many quality papers as possible and stay flexible and helpful to whoever in need.”
Our existence was a defiance of tradition, a diversifier to the mainstream research teams: hyperspecialized and centering around a cult of personality. While these changes should reflect positively in the ideal world, being different is a double-edged sword. Such an experience, while immensely valuable and eye-opening, doesn’t make me favorable in the job market, when I later had to go on a job search. The job market for researchers is bi-modal: you are expected to either stand out in the “first author” pool, or the “last author” pool. You are expected to go as deep as possible when you are junior, and only start broadening when you gain reputation for the depth. I was simply punished for being trained differently.
How it’s going
As you can probably tell, ML Collective (MLC) is basically Deep Collective (DC), our initial team in Uber AI Labs, scaled up to the whole public. All the shaping from DC that had made me different and hence unfit for traditional research hiring offered me a unique vantage point from which to start MLC. The very traits I was punished for when interviewing for regular research jobs were perfectly fit for running MLC (well, to be fair, I did make the job for myself, so it better be a perfect fit).
Even though the organizing and managing practices from DC prepared me for the everyday running of MLC, they were not nearly enough when interest in MLC scaled so fast. I am astonished everyday by the amount of interest from the public, the amount of people wanting to “make it” in ML research, needing a “lab” environment to be empowered, and who are struggling in the process of trying.
What holds a distributed, massively scaled lab together? The answer turns out to be, surprisingly, events. It is the continuity and recurrence of happenings that provide a necessary cadence to produce results, tighten loose ends, create presentations and storylines that gradually turn into a paper. In a formal research organization or any workplace those would be called “meetings,” but, when they are not behind closed doors, we’d like to call them what they are: events.
The two most significant communities of MLC today are almost entirely governed by recurring events. MLC: Lab is the closest to our old DC, where a group of genuinely curious, committed and driven researchers come on their own accord to exchange and execute on ideas, share updates, grind through deadlines, just like in any old fashioned lab. MLC: Open Collab is the massively scaled version of the Lab, asking for a much lighter commitment, and should appeal to those who are still exploring or whose objective is not solely publishing. Our most well-known event, “Deep Learning Classics and Trends”, is a paper reading group we started at my last job and continued to this day. It serves both communities (as well as anyone who needs to read papers to keep up with the unprecedented pace of research), and almost starts to feel like a community of itself.
In a shallow way MLC’s growth can be quantified by the number of members, subscribers, and followers that all these communities gained in a year, but not only is that not the point of this post (selfish takes only, as promised, which I am getting to), it is not even the point of MLC. We don’t really care about it being big; instead we care about it staying active, engaging, and actually helpful. And a lot of times that requires, counterintuitively, staying small.
Honestly, “to be helpful” is the first principle of any non-profit. Although I can’t say how or whether it had been helpful to others for the past year, it has, at the very least, already deeply helped uplift me from a dark period of my life.
All the Selfish Takes
I am recounting all the aspects that MLC has helped change my life for the better. They do not mean to serve as persuasion points for you to join MLC (although you are most welcome to join). If anything, it suggests maybe you should start your own MLC—not a research group, necessarily, but whatever deeply personal community or endeavor your unique position so far has prepared you to birth. But more generally they serve as data samples for you to make informed decisions what’s best for you, e.g. what MLC-like things, big or small, you can take on to improve your life, if you, like me and many others, are feeling stuck, losing purpose, but singularly determined to help yourself out of the rut, by actively taking responsibilities and looking for a change.
I don’t claim to have cracked the code of happiness. My only observation is that it’s never one thing or one moment that flips the switch, but rather an unending journey of trials and errors, one that takes eons to traverse, offers no shortcuts, and rewards only a genuine curiosity to introspect, understand, accept, and build a supportive relationship with yourself.
The Achievement-minded
What do you want to do, and what makes you happy? I find this benign-sounding question quite bewildering. On the surface it seems to be asking the same question twice, but it takes time to realize they are two separate questions, though we easily give the same answer to both without thinking further—for example, “I want to do science; doing science makes me happy.” If we were to answer with absolute honesty, however, you’d have the answer to the first question pointing to the type of work you want to do, and the latter, most likely, to the kinds of achievements you want to obtain.
Piecing both answers together usually reveals something like this:
- I want to change the way people live, and to be known for that.
- I want to make scientific discoveries, and to receive awards for that.
- I want to write books and inspire people, and sell many copies of that.
- …
Too often the second half is left unmentioned; people mix process with outcome, or internal motivation with external validation. Even more dangerous, is when the two pieces are connected by an unsaid “as long as”—when the associated happiness with a chosen endeavor is contingent on a silent assumption, something barely predictable.
To be clear, there’s nothing wrong with wanting achievements or external validation; in fact, I don’t trust any statement of mission without the achievement component. But the fact that people often mesh together these two totally separable aspects, or completely ignore the latter, makes debugging much harder when a problem arises. By laying out the complete statement and separating the process from the outcome, the true desire from the condition, it helps pinpoint the culprit when you are feeling deprived of purpose: is it the work you are doing that has turned out different from your expectation, or is it the outcome? Is achievement a desirable option to you, or an absolute necessity?
A year ago my work was totally centered around achievement. It was hard to notice at first, because on the surface I was just as hardworking as ever, but slowly the initial passion for research was transmuted into a fear of being left behind. Fear bred anxiety, and before long I was a typical achievement-minded: reducing everything I do into a few quantifiable variables, comparing myself against impossible baselines, and worst of all, viewing everyone around me—my teammates, colleagues, people on twitter, people on arXiv papers—as a competitor. While there might be people out there that thrive in competitive environments, I recoil and can hardly be myself in such a mentality.
I work better in collaborative conditions. MLC is created for those that operate like me, who might have once fallen in, but do not ever wish to return to the achievement-obsessed depths. How to create such an environment is actually simple: removing all possible extrinsic rewards, so that one is left to be only motivated intrinsically. We are not your employer or boss, so you don’t get a promotion or bonus from tirelessly collecting proofs of achievement in a document. We are not in charge of your degree, or visa, or career progression, so you are working with us for no other reason than simply wanting to work with us. That said, even if we have the best intention to attract folks with aligned interests, it is still a young and scrappy organization, being refined through trial and error.
The Creativity Axis
I used to believe that I work creatively—after all, I’ve been a researcher all my life, and what is research if not constantly exploring and striving to come up with something new every second? But pretty soon I found myself repeating patterns, chasing easy deliverables, staying in comfort zones, and it struck me that the act of creativity itself—research, ironically, had been conducted most un-creatively. Why is that? I blame the goal-driven narrative that’s so prevalent in the capitalist industry and is seeping into the profession of research. While designed to enhance productivity, it inevitably endues us with the pressure to produce, to quantify impact, and to find the shortest path to ascend. When we enter a new field, we are passed on a consensus as to what is the “right” way to survive; it is an unwritten manual with step-by-step instructions, allowing no divergency from an ordained path.
It may sound like I am ascribing the lack of creativity in the field to nothing but environmental reasons, but no, don’t forget that it is each of us who collectively write that manual. We are willing to write it and adhere to it because the goal-driven style suits us—it’s easy to take in, straightforward to follow, and risk averse in a way that speaks to our covert fear of failure. I’ve grown to learn that it is the insecurity in each of us that’s our greatest enemy to creative living.
I’ve always been insecure, a common trait found in immigrants and in women. Even though I aspire to live an adventurous, creative life, following rules is my innate tendency. At the point of realizing how I had not been performing creatively at work, I immediately felt quite stuck. An easy option would be to look for convenient, conventional changes: change the project, the team, employer, and working environment. Most people change one of those dimensions at a time to stay safe. But while local changes are risk-averse, as any ML practitioners would agree, they usually are not radical enough to push you out of the local minima!
Almost out of pure chance I started running an organization of my own, and it effectively snapped me out of that sticky repetitive pattern. While running any organization is hard enough and requires a lot of brain power, it is especially creativity demanding when it is something no one else has done (I’m not claiming that MLC is the first ever distributed, non-profit, non-employment-based research lab, but it is certainly rare enough that I couldn’t figure out what to do by just “asking around”). We have to come up with new ways of doing even the most mundane things every single day, without so much as a handbook.
I can’t be sure today if running MLC is the thing I was meant to do, but the everyday business certainly brings me closer to what I thought “being a scientist” was like, more than anything else. I’ve got the chance to rethink even the most basic components of an organization: how do we structure, what’s our narrative, how do we recruit, what ties people together, what empowers people to grow. And this eventually feels like the research I’ve always liked: thinking through even the most basic and banal nuts and bolts, over and over, and creating better ways, or even just different ways of challenging the status quo.
But I admit that keeping the creative energy up is only the aspiration, an asymptote that a wiggly curve called reality is forever trying to approach. But simply realizing that clear aspiration gives me a good axis to measure the very metric I care about.
Every day, instead of asking “how am I doing?” which is becoming more and more muddled and indistinct a question, I ask myself its principal component that matters most to me, “how creative am I feeling?”
Whatever my answer is on a scale, I know it precisely reflects a concrete direction in my wellbeing that I ultimately care about.
The Better Connections
It probably goes without saying, but as much democratization as we can get, most of the world’s deals still run by “knowing the right people.” It seems daunting to me, an absolute introvert, that making connections is the major means to make a career advance, or any advance. “Talk to people” at conferences or meetups, used to be the number one advice you’d get entering a new field. I don’t know how easy it might come for other people, but for me it was a constant struggle, especially when I was just starting out. While knowing those “corporate parties” at conferences are important, I never found myself at ease or really enjoyed them.
The techniques required to easily make friends and potential collaborators over “a couple of beers” seem to forever elude me, including but not limited to: having a broad knowledge of almost everything, having a cynical view of almost everything, having a few provocative opinions—either developed on their own or borrowed from someone else, able to keep talking for hours, able to talk over loud noise, able to act or react snarky, funny, dismissive and chippy almost on demand, able to think fast.
I have none of those. At conferences I like to retreat to my hotel room early, order takeout, study algorithms and listen to talks at my own pace.
And I happen to think none of those really speak to the qualities of a good researcher—they do seem crucial, however, to what opportunities a researcher can land.
Despite my best effort I made mostly bland connections during those years. Maybe one out of a hundred turned into something longterm and meaningful, most proved to be a waste of time, and the process was not even pleasant: it was painful trying to blend in and show up as someone I am not.
Instead of general conversation openers at, say, bars, I prefer much more grounded discussions anchored around a concrete topic. A majority of conversations I enjoyed were in front of a poster. But poster sessions have their own perils; for one, it’s either too crowded, where you don’t get any real conversation, or too empty, where it’s too awkward to leave once you realize it’s not something of interest without staying for a good seven minutes for a “quick walk-though.”
The key problem is good research conversions rarely happen, and even when they do, only get you so far in terms of making connections. Two people may quickly click around a research topic, but what happens after a nice chat? A followup email? A formal collaboration? That seems huge and scary and mostly fall through the cracks.
My own experience has been that ever since MLC started I made far more, and far better connections, for three reasons: 1) there’s a central issue a lot of people care about and can get behind: the lack of diversity in the field; 2) there’s a solid solution framework we are actively building; 3) there are concrete asks, actual ways one can get involved, no matter how senior you are or what levels of commitment you are comfortable with.
Also, by “better” connections I don’t mean just bigger names, but the “right” ones that deeply care about the same issues as I do, have thought along the same direction as I have, and can brainstorm ways that truly help make our mission stronger in one way or another.
Better yet, such connection-building is totally introvert friendly. There’s not a lot of grandiose, off-the-cuff talking, but more offline writing, rehearsed presentations, and if you are really on the far end of socialization avoiding like I am every so often, you can drop an RFP and just silently wait to be contacted with plots and reply only if they are beautiful.
Before I am Ready
There’s a chicken-and-egg problem everywhere: to be given the opportunity to do something you have to prove that you have done it in the first place. To be hired as researcher you have to have already published papers. To be given a chance to manage you have to have had direct reports before. But how does the very first 0 to 1 change happen?
It usually happens in one of the following forms:
- Pure luck. There’s a vacancy, a shortage of candidates, there’s a crisis or a breakthrough and you happen to be in the right place at the right time.
- Some luck plus a lot of trust. There’s an opportunity but also you’ve built a strong trusting relationship with the right people for a long time.
- Pure trust. All you’ve done is building a strong trusting relationship with the right people for a long time.
Any form of luck is, by definition, rare. And any form of trust requires time. What if you don’t have either? My best lesson had been: do it before anyone tells you to do it, before earning any so-called legitimacy. Prepare yourself well enough when opportunity knocks on your door.
A year ago no one was willing to hire me to mange or lead any size of a team, because I lacked the formal track record of leading (even though I had been de-facto leading); it woefully stings the same way the rule of research does: no accepted paper, no proof of research. I segued to growing community and running an organization for a year, and have since factually upgraded myself in the hireability food chain.
In the same way, MLC is a place where you don’t need to be given a researcher title to start doing research. It may serve as your “preparation” or “crafting skills” period before eventually landing a researcher’s job, but I don’t suppose that to be the sole purpose. I wish both idealists and realists would enjoy having a space and community to do the work they are not told to do, for whatever purpose there might be.
A More Sustainable View of Philanthropy
If you are like me, you probably have heard or read about the “giving makes you happy” spiel many times before you were ready to become a giver. And if you are like me, you probably keep hearing it from the richest and most successful people of every profession. As a result, before I learned to internalize the message I first learned to dismiss it. “It’s easy for you to say. You are a billionaire.” Another proof that when sampling is not diverse, the method is not effective.
I used to think, yes, I am a kind, giving person, but to start becoming a philanthropist I have to first make it somehow. The notion of “making it” vaguely equates to some kind of title, recognition, fame and reputation. I couldn’t define it concretely at the moment, but I’d surely know when I get there.
But that’s just a big illusion. We, of course, will never really feel like we’ve “made it.”
My biggest learning this year had been that giving does boost your happiness to a whole new level, no matter what status of life you think you are in. And giving and taking are never mutually exclusive; in a way, giving is taking. I wish someone had told me earlier. But the strange tragedy is that they had. Only the thousands of messages I received early in life were never delivered in a way that I could relate, that would have pierced my defenses, that I could have absorbed like a sponge, instead of rediscovering it through the hassle and heartbreak of life, as if those struggles were strictly prerequisite for its attainment.
From a macro level, if only the most wealthy and famous are doing the giving, if it’s an activity that one’s only option is in (1) or out (0), philanthropy is not sustainable. MLC bears my renewed belief that we build a better world when the act of philanthropy can be adopted in a fractionalized way. If you are really well off, your philanthropical involvement could be close to 1. Otherwise, anywhere on the axis between 0 and 1 should be viable. In a more traditional form of giving, that could almost exactly map to the amount of money one donates. But in the ML research world, the right currency has yet to be defined. Before that, most giving activities carried out by the most prominent figures are still seen as a binary option: you are either all-in to take in a mentee or commit to collaborating on a project, or you are still waiting to feel ready to do it.
MLC wants to create a market where you can contribute “fractional” goodwills. You can donate one hour, one idea, in a one-off fashion by attending any one of our recurring events, all the way to a committed X hours per year by joining as a researcher. We also believe the “giving is taking” concept is even more true in ML research mentorship. Ask any professor who’s had students, and they will tell you how much of their advising is intrinsically inseparable from directly learning from their students.
One lingering concern I have when writing this section is that there’s no way this message can reach those that were like me a decade ago, when I didn’t believe that I had any link to the grandiose notion of philanthropy because I was poor and only a researcher-in-training, having no idea how much she can already gain by starting to give.
But there’s still hope, that there is somehow a serendipitous way this message could reach them. At the very least, even if I am just preaching to the choir, perhaps it’s not entirely worthless, as the brilliant Maya Angelou once said,
“To write about giving to a person who is naturally generous reminds me of a preacher passionately preaching to the already committed choir. I am encouraged to write on because I remember that from time to time, the choir does need to be uplifted and thanked for its commitment.”
What are MLC-like things to you?
There are a lot of pedagogical posts out there trying to be your teacher—ones that speak with a lot of confidence and authority and deliver messages in a form of “You should do X.” I don’t want to be another presumptuous teacher. My favorite scientific papers, and the type that I try to write, are usually written in this tone: “We observed X. We found it useful. Maybe you should try X in your experiments too! If you do, let us know what you find.” In the same spirit, I hereby present observations of what made my life better in the past year. If the problem you are trying to solve falls into the same condition as mine was in, maybe you should try them too. If you do, and are willing to write down your observations and pass them on, a small progress is made in the collaborative project of “self-help in ML research.” Together, we may one day converge to something closer to the truth.
- Own a thing—one that you feel comfortable nurturing. A garden, if you may. But better if that thing involves people. After all, an invention has no footing if it doesn’t ultimately serve people.
- Find an interest that’s centered around not a problem, but a meta-problem. For example, instead of thinking about a benchmark to beat in ML research, think about how ML research is being done, how researchers are doing, whether there are things we can do to make it better.
- Find a place where you feel a sense of belonging. Maybe it’s your team at work, maybe it’s someone else’s team that you are comfortable hanging out with, maybe a friends group, a workshop, a club. After finding it, think of small ways you can contribute to making others feel they belong too.
- Find a task that allows you to stay creative. Don’t be fooled by whether the definition of task contains the wording “creative” or not, observe yourself to gauge if you are actually operating in the creative zone.
- Don’t accept the status quo, change it, if you see it less than perfect (believe me, it never is). Even if just a little bit.
- Give. Try a small amount first if you are unsure. It is truly the best way to take.
Thanks to Joel Lehman, Katherine Lee, Ken Kao and Laurent Dinh for generous edits, feedback and discussions throughout the making of this post.