Feb 3, 2017 · 27m · mad
Matching People to Careers Bias-Free // Frida Polli, Pymetrics (FirstMark's Data Driven)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Data Driven NYC presentation, Pymetrics co-founder and CEO Frida Polli demonstrates how combining cognitive neuroscience games with machine learning creates a bias-free candidate screening process. She explains how objective behavioral data replaces flawed resume reviews to optimize early-career hiring, boost employee retention, and foster workforce diversity.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 2.6% of the talking time here. How this is scored →
Matt as informed peer 0.8 Guest teaching 4.1 Guest disagreement 1.4 Matt pushing back 0.5
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 24:00 Defending against workforce homogenization claims
Frida forcefully rejects the audience member's premise that data-driven matching creates a homogenized workforce, arguing six-second resume scans are far more subjective and homogenizing.
Hardest push from Matt ▶ 21:50 Challenging feedback loop reliabilityMatt questions the premise of Pymetrics' feedback loops, asking how they separate true candidate fit from noisy performance data and specific manager biases.
Biggest teaching moment ▶ 13:05 Flawed contrast groups and the NBA analogyFrida educates the room on statistical contrast group design, using an NBA player metaphor to demonstrate why contrasting top and bottom performers yields misleading features.
Matt holds his own ▶ 21:50 Drilling into data reliability and feedback loopsMatt demonstrates clear technical understanding of product feedback loops by pushing Frida on how Pymetrics isolates employer performance signals from individual manager noise.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Audience Engagement and Overview of Career Matching | 0 | 3 | 1 | 0 | Frida delivers an opening presentation explaining the core recruitment matching problem and inefficiency of resume scanning. The host does not speak during this monologue segment. | |
| Neuroscience and Data Science Foundations of Pymetrics | 0 | 4 | 1 | 0 | Frida explains how Pymetrics leverages neuroscience research and game-based cognitive assessments rather than traditional questionnaires. As a monologue presentation, the host scores remain zero. | |
| Comparing Traditional Subjective Data to High-Resolution Behavioral Data | 0 | 4 | 1 | 0 | Frida details why resume data is low-resolution and subjective using a facial recognition software analogy. The segment is entirely guest presentation monologue. | |
| Demonstration of Pymetrics Game and Trait Metrics | 0 | 4 | 1 | 0 | Frida demonstrates the Tower game and outlines standard machine learning techniques like regularization and cross-validation used to prevent overfitting. The segment is a monologue presentation. | |
| Defining Contrast Groups and Semi-Supervised Learning | 0 | 5 | 1 | 0 | Frida uses an NBA player analogy to critique traditional top-versus-bottom contrast groups and explain baseline selection in semi-supervised learning. The host is silent during this presentation segment. | |
| Automated Processing, Trait Reporting, and Bias Auditing | 0 | 4 | 1 | 0 | Frida explains fit scoring across multiple roles and contrasts Pymetrics with 80-year-old assessment instruments like Myers-Briggs. This is a monologue presentation segment. | |
| Measuring Impact: Reduced Application Times and Candidate Marketplace | 0 | 4 | 1 | 0 | Frida outlines enterprise case study results and presents interesting cross-career trait correlations comparing hedge fund managers with venture capitalists. The host only offers a brief closing joke as Frida finishes. | |
| Q&A: Model Feedback Loops and Target Hiring Demographics | 6 | 5 | 4 | 4 | Matt opens Q&A with a sharp question probing how Pymetrics validates feedback loop data against manager-level noise. Frida also defends her algorithms passionately when an audience member asks if the tool creates workforce homogenization. |
Statements from this episode (15)
The average recruiter receives 250 applicants per job posting
“The average recruiter gets 250 applicants for every job that they post.”
Recruiters manually scan resumes for an average of six seconds
“If it's a person doing the scanning, they'll on average scan it for six seconds”
High-density objective data yields superior prediction of human outcomes
“And so what we just, what we learned in our, you know, 10 years at Harvard and MIT doing brain imaging is that if you collect better, more objective, more high density data on people, you can actually do a better job Of predicting research outcomes.”
Nonverbal neuroscience games avoid cultural biases found in questionnaires
“And we assess 90 different cognitive, emotional, and personality traits, and again, these are all nonverbal, ah, games, so they're not culturally anchored the way, you know, questionnaires or other things might be.”
Neuroscience games lack inherent gender or ethnic bias
“Unlike some of the other data that people are using to predict fit in people, the games were generally, are generally not thought to have any inherent gender or ethnic bias, so they're not picking up on that if the training sample is biased.”
Resumes and questionnaires are subjective, low-density hiring tools
“Both of these are low density and also subjective.”
Process metrics provide better algorithmic features than binary pass-fail outcomes
“This information is a lot more useful in terms of, ah, building algorithms for, as features for building algorithms than just sort of binary, binary outputs.”
Most recruitment startups simply apply machine learning to resumes
“So there are a handful of startups that are starting to do this type of work in the recruiting field. Many of them are focusing on the resume as a field, as a source of data.”
Candidate models require a broad baseline contrast group
“What we're saying is, no, you really need to collect a baseline that is representative of the people that you're, you know, trying to select from and use that as your contrast group.”
The most commonly used career instrument, Myers-Briggs, is 80 years old
“The most commonly used instrument is the Myers-Briggs and it is almost a century old at this point. It's 80 years old.”
Behavioral data shows VCs chase high rewards while hedge fund managers hedge
“VCs are very motivated by that sort of high reward stuff. Versus the head fund managers, they're hedging, right? So of course they're going to put in effort for a variety of different things, including low reward situations.”
Predictive hiring tools are best suited for early-career candidates
“Yeah, so I think that this solution is honestly best suited for, let's say, the first 10, 15 years of this process. I don't think it definitely can work with experienced hires, but the reason I think it works better earlier on is a couple reasons. That's reall…”
Assessments may measure 90 traits, but only 10 determine job fit
“The second thing I would say to that is we measure 90 different traits, right? Of the things that are end up being important, it's usually maybe 10. Maybe, maybe fewer than that. The remaining 80 are left to vary tremendously, right?”