Eran Arkin
4y ago
Visit me to learn about Tech Careers, Product Management More. Experienced operator; Beginner online writer. Dad.
I'll walk you through how AI product management differs from traditional product management based on my experience working on traditional and ML based projects at Google, Waze & Meta.
If you haven't yet worked on a problem requiring ML, you will likely encounter one in the next decade as “Software Is Eating the World, but AI Is Going to Eat Software”.
Making the transition isn't hard as it may sound. Let's dive in.
Online Data is Not Only the Output. It's a Key Part of the Input
Savvy product managers proactively utilize existing sources of data and find new ones. Think about how TikTok is designed: because there's exactly one video on the screen, the program knows exactly how long you watched and what actions you did on the video. This enables it to recommend the next video almost perfectly.
You Don't Control the Recipe, Just the Outcome
In traditional software, the user inputs, and the product team determines the exact steps to generate the outputs. In ML, you define the desired outcome, specific performance criteria to be optimized, and the program determines the steps. The beauty of it is that the program can discover some unobvious ways to solve problems [watch this video!], but....
Be Careful How You Set the Desired Outcome
Let's say you run a content recommendation feed and optimize for clicks. Your engineers launch a new model, clicks go up, and everyone gives high fives. You look at the top content dashboard, and suddenly you'll see a lot of bad stuff, clickbaity, low-value content, soft porn, etc. You didn't create the product experience or feeling you wanted, and advertisers may boycott you. Or you might push users too aggressively to harmful rabbit holes.
Junk food is immensely popular & addicting, but it might not be the restaurant you want to create.
Your XFN is Larger
To name a few common additions:
Specialized ML Engineers to build models.
Project managers & ops people to manage offline rating processes.
Privacy experts: you'll want to retain data longer since data is fueling model performance
You Don't Need to Know the Ins & Outs, But Understanding ML Concepts is Important
There are less than 100 car manufacturers in the US and more than 200M drivers. Drivers don't know how to create a car, but they know how to drive & read the dashboard.
Some key things you'll want to know (DM if you want a post on the best resources) before you are ready to drive -
Supervised / Unsupervised learning
Optimization & Loss function
Precision, Recall, False Positive, True Positive, False Negative
While There are few differences outlined above, the most Important PM skills, are the same: critical thinking, communication, and user empathy
Good luck in your ML Journey. Drop me a note or DM if you have any feedback. Thank you for reading