Last week, we published an interview with Camile Fournier, ex-CTO of Rent the Runway, in which we talked about managing technical teams, multidisciplinary teams, and the struggle to develop good managers in product teams. This week, our founder, Francisco Homem de Mello, interviews Jeremie Harris, founder of the Canadian startup SharpestMinds, which aims to find the best way to hire data scientists at the start of their carreers.
SharpestMinds works as a “blind” recruitment process, in which more value is given to the knowledge and skills of the data scientist, as opposed to their credentials, such as education, grades or previous experiences.
But the value of the company goes beyond the recruiting process: the program is filled with challenges and projects that the candidate completes with the company, in a journey that improves the chances – and the skills – of each candidate.
Kiko: Tell us a bit about what made you found Sharpest Minds, and what’s your vision for the company…
JH: We noticed that companies that lead the AI pack treat recruitment very differently from those that don’t. The top dogs in data science and machine learning – the likes of Facebook, Google, Airbnb and Amazon – all have very active campus recruitment operations.
That’s their secret weapon: unlike companies that struggle to find high-performing talent in machine learning, they’re able to catch new talent quickly, scooping up new grads (and even some current students) through internships and work terms.
It’s a huge advantage, and there are two reasons they can do it, and others can’t: first, they have the ability to reach tons of students through their relationships with labs and schools, and second, they have enough pre-existing deep learning or data science expertise to be confident in their candidate screening procedures.
Our vision is to let exciting startups that haven’t yet reached Google scale compete with Google to access top talent before anyone else. We think access to talent needs to be democratized, and a huge part of doing that is providing a single robust and highly effective screening process that any company can lean on as it establishes or grows its machine learning team.
Kiko: What’s the difference between hiring a data scientist and hiring a software engineer?
JH: One nice thing about data scientists is that there are very objective ways to assess their performance. Benchmark datasets are great because they allow you to determine that someone is in the top X% of their peers when it comes to a particular task.
But another key difference is that data science is, by nature, more open-ended than software development. You can actually hand someone a dataset and ask them to provide you with all the insights they can from it, without being any more specific than that. The ability to come up with new ways to extract value from the same dataset is a key attribute to look for in a data scientist since there’s often more value hiding in your data than you might realize.
Kiko: What makes for the best data science candidates? How important are credentials x skills x experience?
JH: Credentials and experience are proxies for skill. We don’t care about them at all. Our entire process is background-agnostic, in the sense that we don’t worry about whether someone is an MIT PhD or a high school dropout as long as they’re great at machine learning (and specifically deep learning), and know their way around a dataset.
We’re obviously biased, since our team consists mostly of physicist dropouts who taught themselves deep learning. But in many ways, that’s given us insight that many companies don’t have. We firmly believe that with the resources that are available out there, anyone can teach themselves how to be effective machine learning developers. It’s not easy to do, but getting yourself industry ready on your own is definitely possible – everything you need is open-source – the tools, the papers, and even compute is increasingly getting accessibly cheap (or even free).
Kiko: For a software engineer, it’s easy to find top-of-funnel talent for a selection process: just go for computer science and computer engineering programs. But there isn’t, that I know, a data science degree. Where are these candidates coming from? Math? Statistics?
JH: A lot of the best machine learning engineers and data scientists are cross-overs from other fields. Although there are some data science degrees popping up these days (for example there’s a great push in this direction by the Canadian Vector Institute, and other academic institutions like Berkeley), I’d say the majority of top candidates are still coming from these varied and unorthodox backgrounds.
We definitely see a lot of physicists, statisticians and mathematicians who have become great machine learning developers. Their main challenge usually ends up being on the devops side – they have to learn the ins and outs of software development to be able to deploy scalable models and play with large datasets.
On the other hand, engineers and CS majors usually have the devops down, and need to spend more time learning the math behind deep learning and classical machine learning algorithms.
So it’s definitely a mixed bag. Far more people are capable of being high performers in this space than most realize (part of the reason why we’re such huge fans of what’s being done at Fast.ai).
Kiko: Your program takes a bit of time, during which you prepare candidates for the “real world”. Is there a disconnect between the output of top colleges and the needs of companies?
JH: Colleges and universities are by and large pumping out people who can write papers but not always clean, functional code. And that makes perfect sense – professors face the classic “publish or perish” incentives of academia, and encourage their students to think the same way. They’re rarely as interested in solving real-world problems as they are in getting their work published in Nature or NIPS.
Fortunately, many students understand this and take steps to ensure that they can write readable, efficient, production code. But they’re not learning that in school for the most part.
Kiko: We all know how hot the market is for top talent. And I know that SharpestMinds has a stringent selection process, that aims at filtering the top 5% of candidates. What makes a candidate not take an interview with, let’s say, Google (path of least resistance) and go through the Sharpest Minds program instead (more resistance)? In other words, how does anyone implementing a stringent selection process makes sure he’s not adversely selecting the worst, most “desperate”, candidates?
JH: The short answer is, you need to make it more appealing for people to come through you than through other options.
New grads and students who become SharpestMinds members have three huge advantages over everyone else.
First, they get screened and interviewed by us, which means companies don’t generally make them go through all their standard hiring/evaluation steps. Our candidates usually go straight to final interviews with the companies that hire them. That’s a big savings of time for the companies, but it’s even better for the students, since a single round of interviews might get you through the door at 3 to 10 companies.
Second, they can start working before they graduate. Our focus on discovering talent early extends to connecting businesses with students before they’ve completed their degrees. Our infrastructure can accommodate in-person, remote, part-time or full-time engagements, which is what you need if you’re serious about reaching Googlers before Google.
Third, they’re part of a broader ecosystem of developers and new grads that provides them with exclusive access to the resources they need to succeed.
There are clear indications that we’re effectively selecting for top-tier candidates too: our last cohort of applicants featured former interns at Google, Apple, NASA and Tesla Motors. So we’re catching the conventional cream of the crop, but we’re also finding that there’s a surprising amount of high-potential undiscovered talent out there that everyone misses because they’re blinded by a perceived need for prior experience (again, is a proxy for what they actually want, which is actual skill).