Why Can’t We Find Fulfilling Work?
How machine learning can help us build a better future by making it easier for people and jobs to find each other.
At our last company, my cofounder Josh and I were building a state-of-the-art laser raster VR display and needed optical engineers. We put up a job posting and waited. Nobody applied.
Desperate, we scoured the web for optical engineers, writing personal, thoughtful messages explaining why we thought they’d be interested. This proved successful — responses poured in, and we ended up working with the lead optical engineers from Google Glass and Hololens.
But in doing this ourselves, we thought: this is insane. How could it be that in 2016, people still learn about great opportunities via random chance — because somebody found them on the internet and cold emailed them?
How could it be that in 2016, people still learn about great opportunities via random chance — because somebody found them on the internet and cold emailed them?
It’s not easy to find your best opportunity —
Because what you’re looking for can’t be captured in simple search keywords.
In a job, you look for things like “offers me opportunities to advance” or “helps me get better at Javascript”. And in a candidate, companies look for things like “wants to grow their people-management skills” or “scaled the backend at a company with millions of users”.
It’s hard to search for these attributes on LinkedIn or Google, because they’re complex, multi-dimensional concepts. That’s part of why so much time is wasted on both sides finding the right match — and why people remain dissatisfied in their jobs for years.
A new approach: a machine learning model for every job and human attribute.
Turns out, machine learning models are especially good at capturing the underlying traits of high-dimensional concepts. At Sourceress, we’re applying this to make it possible to search for the complex ideas.
Working with companies, we create a custom-tailored model for each role to capture the nuances that job descriptions inevitably brush over. (You might say you want 5 years of Android experience, but would you take 4 for someone who’s been programming since childhood?)
On the candidate side, we’re creating a skill and interest taxonomy that goes far beyond tallying years of experience with Python. We’re building thousands of models that run over training data labeled by our subject matter experts to quantify each attribute. This allows us to calculate the probability that a person is a good match for a role.
This approach is working incredibly well: we take over the entire sourcing process for companies, and when we engage candidates, hiring managers are excited to talk to them more than 85% of the time! (Compared with 10–50% for the typical external recruiter.)
As a result of our high calibration accuracy, we’ve made many hires for our customers, grew revenue more than tenfold during Y Combinator to over $1M in annual run-rate, and are excited to announce our $3.5M raise from Lightspeed Venture Partners and researchers at OpenAI.
40% of our hires for customers are female engineers.
We want to be intentional about using this approach to reduce bias and increase diverse hires (beyond gender diversity). Here’s what we’ve learned so far:
- Our models are less biased than humans, so we qualify more diverse candidates. During human quality control, we hide names and photos. I’ve been shocked at my own internalized biases: more than once, I’ve sat in on human quality control for an amazing candidate, only to feel surprised when I later learned that the candidate was a woman.
- Underrepresented candidates are more likely to disqualify themselves when they see a job description. But when we tap them on the shoulder and reach out in a thoughtful way, they’re more likely to respond than non-minorities.
- Underrepresented candidates are less likely to highlight quantifiable results on their profiles — so we look for evidence of skills that hiring managers are weighting strongly, even if the candidate doesn’t mention them explicitly.
While we think machine learning has incredible potential to decrease bias in hiring, models can very easily reinforce bias if we’re not careful. Preventing bias is something we’re actively working on, and we’ll blog about it in the future as we learn more.
The path to better employing humanity’s intelligence.
In the short term, we aim to reduce bias in the hiring process. In the long term, we want to remove the barriers to finding fulfilling work.
Add up all the knowledge workers in America who go through the motions for 8 hours a day — and you get millions of hours of wasted thought work every year. If each person were instead in their perfect job… how much more progress could we make?
By creating models for people and jobs along these thousands of dimensions, we’ll be able to show candidates the best jobs that match their skills and interests. Which means that even while you’re employed, you’ll know what positions you’re qualified for, where the hiring manager would talk to you immediately. No more sending resumes into the void.
We’ll also understand enough about open opportunities and the skills required to help you navigate your career. We could suggest next steps for gaining the skills and credentials that unlock the places you want to go. And maybe even create the opportunities or education programs that will get you there.
If we can reduce friction to finding higher impact work, we’ll help people be more productive, feel more fulfilled, and ultimately accelerate human progress.
Want to help people work in jobs that matter?
Every engineer here has the opportunity to build their own models to quantify people’s skills, interests, and character traits. Whether you’re interested in machine learning or whether you enjoy iterating on product with customers, there’s a ton of opportunity to build something that truly reduces friction to finding fulfilling work.
We’re actively hiring exceptional founding team members in engineering, client relations, and operations. Check out our jobs page and values to learn more!
