Starting a Ph.D. level career in AI/ML

Introduction

I was recently asked: What advice would you have for someone in Artificial Intelligence / Machine Learning / Data Science to have a successful career in industry, if they are just getting started post-Ph.D.?

Since I’ve been a manager at tech companies for 15 years, I’ve been asked similar questions a lot. I thought I’d put some answers down. Pardon in advance if they are broad and sometimes cliched answers – I think they still work. Were I mentoring someone in an actual 1:1 we would pick a narrower topic, go into more depth, and debug how to go about some piece of this.

The opening parts lean toward applied ML careers, but after that first section I talk about what in my experience has seemed to help engineering careers broadly.

Prepare a history of impact for job interviews

Establish a track record of impact before you graduate. Unless you are intending on pursuing an academic career, publishing articles without applied/product impact has a diminishing return, after some volume. Impact includes original scientific invention (if applicable to a contemporary problem), moving significant business metrics, or landing significant product features. Open source productivity is welcome but should pass a significance threshold.

Can you take an abstract or ill-defined problem, negotiate with stakeholders to turn it into a set of metrics to move and constraints to hold fixed? Are you rugged and diverse in the range of methods you can apply to move those metrics? If you are expert in just one niche: either you need to find a business passionate about that niche, or you need to diversify. Can you handle problems at distributed scale – or has academia prepared you only for toy problems? Can you deal with the vast messiness of real-world problems? Are you appropriately skeptical of data? It's likely polluted or riddled with inconsistencies, your ground truth may be wishful thinking, the prior generation you inherited may be dotted with methodological errors. Anyone who has been stewing too long in any given problem has probably found a way to make their training set pollute their eval set. In any case, an engineer who thinks entrepreneurially owns the outcome and learns whatever it takes.

Since ML for images and large language models has shown such amazing utility of late it can be tempting to use and trust in deep neural nets somewhat blindly. But an investment in learning a specific domain's knowledge with some depth can really pay off. Talk to the experts in the given use case, and get your hands dirty understanding the data.

Be careful about jumping at the first opportunity

Being new in your career, take some time to get perspective and look around at a variety of opportunities before you make a long term commitment. If every job you accept turns into a 3-4 year commitment, you've got 10-15 shots. Make them pay off! It's easy to blow a decade on unfruitful directions. If you can, see who you know at a company. If possible get some back-channel opinions. When I’ve made major career decisions I’ve generally literally put the options down the rows, and the factors across the columns of a spreadsheet, written down the facts and wrestled with the weights. I’ve used a couple of trusted advisors as sounding boards.

Personality

Demonstrate your ability to work as part of a multidisciplinary team. Communication and teamwork skills are just as important as any technical skills. Be a servant to the team's success. Be a force-multiplier for others' careers and achievements. They will return the favor.

When I was with Google, there was a “No Jerks” manifesto circulated. Be humble; pride impedes growth. Posing can never hide failure. And if you don’t fail sometimes, you aren’t measuring yourself against truly worthy challenges – so don’t be afraid to stumble but document your missteps and write a good retrospective. When I was at Facebook, “Feedback is a gift” was oft-repeated.

It can be cliche but some engineers enter the workforce with awesome technical skills, but weak interpersonal and social skills. If that fits you at all, work on it. Improve your language skills, conquer having a strong accent in English or your host-country's language. Join a speech club (like Toastmasters). Take acting or comedy improv classes. Learn social dancing. Fully engage with nontechnical people -- and expect that the longer your career runs the more cross-functional your scope is likely to become, so get practiced at working with all modes and personalities.

Direction

Have a 5-10 year plan for your career. Hack your life. As a software engineer, you know the power of structuring problems, and recursively solving subproblems as preconditions. Do that too for your own career development. Select roles that create the growth opportunities that will make you the person you want to be. Pick role models who resonate with you, and pursue a montage of traits and abilities you can make your own. Be intentional. A career lived by Brownian motion may not take you as far or as fast.

Seek appropriate risk. That is, wise “alpha bets” that have a high return on skills gained, people met and professional relationships developed, or just plain equity/comp which can set you free. In the vibrant technology industry, getting to financial independence has been very plausible for many people. Will it be so in the future? No crystal balls here, but I see AI/ML as still booming. There may be moments when you cannot so readily take risks: family commitments, health challenges, etc. So make the most of opportunities while you are unencumbered. (But … being encumbered can be the best part of a meaningful life – narrow workaholics can be sad.)

Reinvent yourself to some degree every 3-4 years: you’ll likely have hit the flat spot on an S-curve of growth, and could benefit from an adjustment. That could mean switching projects, technologies, or companies. But if you’ve got a vehicle of success, ride it and “innovate in place.” Don’t be so restless as to walk away from a sweet spot. On the other hand, even if you are happy at an employer, it can be worth taking a span of time to do a job-hunt every 3-4 years. See what people are talking about in the industry. See if there are some burgeoning areas outside the sheltered walls of your company. Keep in touch with entrepreneurial-minded technical people. You may put clues together and be able to take a great leap forward in opportunities.

Conversely, as an engineering manager I often perceive people as having implicitly chosen to put a cap on their careers. It becomes impossible to promote them past, say, SWE-IV (to pick something arbitrary), because they have defined themselves as a person who does just XYZ. The other things are not in their job description. The new skill, the exposure to risk, the certain experience of awkward fumbling and failure needed to move on, is something they have decided to not assume. I hope you break out of self-imposed limits: the job description is – solve the real world need, delight the customer, protect humans from harm, etc., do anything it takes irrespective of technical and role boundaries. Your opportunity to serve is limited only by how much responsibility you can shoulder.

Specifically, the "default data scientist" (not there is such a person), hasn't developed: project management, people management, low level coding, GUI or systems architecture skills -- just to put out a few potential examples. Rather they've focused on moving and analyzing data. Learning your adjacencies and having a broad toolkit gives you options. A reputation of breadth added to your in-depth knowledge makes you a person likely to be drafted into leadership.

Passion

As an interviewer, I look for: where does this person exhibit passion? What feeds their energy, and drives them? Engineering and science can be hard, frustrating, and sometimes gruelingly monotonous until you break through and complete tasks. So become the kind of person who has the motivation to endure those hard phases of work, to sustain and lift up yourself and others on a team who may be flagging.

Know yourself. Broad, full-stack, core tech or applications, core algorithms, infra, or operations (ML Ops) – where is your passion? People sometimes think the core algorithms are the best or only place to be, but you can thrive anywhere if you find business needs then discover how to delight in solving the problems, and own the solution. Enjoying your work is a kind of superpower.

Stay curious. The best engineers I know have side projects they pursue - because it brings them joy. Joyful optimists may not always be realistic, but they win disproportionately – and have more fun doing it. Being happy and healthy are career enhancing, aside from being frankly more important than just a career.

Continuous learning

Try to find mentors who are notably ahead of you in your career path, as well as who contrast with your path (for diverse perspectives.) A good mentor is generally outside your change of command, has objectivity and no skin in your particular game. (In contrast, your chain-of-command has an inherent conflict of interest, as they want you to deliver what’s best for them and the company, and while that is often aligned with what’s best for you, sometimes it’s not.) You can ask your manager to help you find a mentor; there are also out-of-company mentors.

I appreciate the concept from Google of “20% projects” (although I suppose the actual practice dwindled and faded with time.) But the concept is that you should dedicate a portion of effort toward projects or role-changes that drive learning and push your personal edge. Stephen Covey referred to these sorts of activities as “sharpening the saw.” If you are new to industry, well you spent a lifetime gearing for this moment. It’s time to do! To have an impact. But if all you do is land impact and omit the self-care of continuous learning, you’ll find yourself in a rut.

As a data scientist, there is one advantage that you might have (depending on the company). Can you get into a role where you have access to data that cross cuts an entire enterprise? Can you find a role where you can serve many different clients inside the company over time? Last, can you produce data that rolls up to reports to senior leadership, and so gets you a seat at the table? This may not apply to all, or even many, data science roles. However, I observed that at Facebook (now Meta) the Core Data Science (or CDS) team, when I was there had all of the above cross-cutting potential. Such exposure broadens your vision of what's possible, and then perhaps you gain the inspiration and openings to choose more consequential projects. While I have not done it, similar opportunities can arise outside giant companies but working as a serial consultant.

Post Ph.D. specifically?

Don’t think too highly of yourself because of your Ph.D.; people who learned in the school of hard knocks may well outperform you or soar past you. Especially, the differentiation of a Ph.D. washes out after some number of years. The real test is: are you continually growing – and reifying that growth into artifacts that demonstrate impact, creativity and service?

Obtaining a Ph.D. is a prequalifying step. It should have forced you to gain deep expertise in at least one narrow domain. You should have grasped the basic elements of computer science’s major branches. You should know the scientific method, in the particularities with which it applies to your discipline. You should know the basic outlines of the contemporary research landscape. For example, you can outline all the major branches of applied machine learning, know the common pragmatics of applying it, and can rattle off the more significant innovations relevant to your discipline in the last several years. It’s forced you to communicate your thoughts clearly and professionally. So you have a leg up vs. everyone who has neglected to take on the pains and suffering needed to establish those skills.

If your particular discipline leads you into your distinctive career niche, and you still enjoy it, awesome! Run with it. And for those in AI/ML backgrounds, that is typically very feasible. The industry is bursting with need for applied skills. However, Ph.D.s can do perfectly fine migrating between technical disciplines. Your degree is no straitjacket. People change their majors all the time, even late in their careers. Non-software and non-data science Ph.d. often adapt quite well into C.S. disciplines (given some retooling). For example, physicists can often outperform. They have the mathematical rigor, and they approach complex dynamic systems with comprehensive thinking, a wariness to ferret out errant assumptions and the understanding that they need to uncover the structures that govern the particular technical domain.

Closing

I've had the privilege of partnering with supporting many hundreds of applied AI/ML engineers and data scientists, across five companies now. Often the companies I've served at have had surveys where people have been asked: what do they enjoy most about working at (Company X). The most typical answer has been some form of "I most enjoy working with amazing team mates." So, my closing wish is: have a terrific career in AI/ML/DS with a community of fantastically creative, talented and dedicated people.