Humans + Algorithms:
Who Decides?

Adriel Frederick

VP Product, Reddit X · ex-Lyft Marketplace · ex-Facebook Growth

LENNY'S PODCAST

Core Framework

The Human–Algorithm Decision Divide

"Your job is figuring out what the algorithm should be responsible for, what people are responsible for, and the framework for making decisions."

Algorithms excel at optimization within defined constraints
Humans own strategic judgment, intent, and long-term effects
The PM's role: design the handoff between the two
Leaving everything to an algorithm is a product design failure

The Lyft Marketplace Lesson

Operational Control Is a First-Order Product Requirement

"You have to think about operational requirements and operational control as a first order requirement — we got caught up in the algorithmic complexity and sweet sauce of it."

Lyft operated across 300 cities — no two markets identical
Algorithms couldn't see snowstorms, competitor moves, or new taxes
First build: no human control levers → marketplace couldn't be managed
Fix: rebuild with people explicitly in the loop alongside the algorithm

The Snowstorm Problem

A snowstorm hits Chicago. Driver supply collapses. Your pricing algorithm has no context. A human needs to see the situation, decide the right incentive level, and push it through a product interface — that interface is a PM's responsibility to design.

The Race-to-Zero Problem

Tell an algorithm to maximize market share. It drops prices to zero. Add a profit floor. Both you and your competitor hit the floor simultaneously. The game stops. Now a human has to decide where you want to win — that's a judgment call no algorithm can make.

Adriel's reframe ML is a tool like a screwdriver. PMs decide how much to put in the tool and how much to leave to the person holding it. Give them a flathead, a Phillips, a Torx — let them pick.

Playbook

Building Algorithmic Products That Keep Humans in the Loop

Step 1 — Map the decision space: List every judgment call the product makes. Explicitly assign each to human or machine.
Step 2 — Design the ops interface first: What information does a human operator need? What controls must they have? Build this before you optimize the algorithm.
Step 3 — Bound the algorithm: Define hard constraints the algorithm cannot cross — price floors, safety rails, ethical limits. These are human calls baked into the system.
Step 4 — Surface strategic context: Give operators visibility into market conditions, competitor moves, and goal performance that the algorithm cannot infer from data alone.
Step 5 — Close the loop: Let operators push intent into the algorithm in real time. Amplify human judgment rather than replace it.

Facebook's Equivalent

Even at Facebook's scale, someone had to decide how often to show ads vs. organic stories vs. discovery content vs. friend-finding prompts. Algorithms optimized within that choice — but the strategic allocation was always a human call, varying by market and moment.

R&D Team Survival Rules

Innovation labs die from organ rejection. Stay alive by: (1) tying your work to the core mission, (2) making wins shared across the org, and (3) ensuring other teams still feel empowered to innovate — you're not the only source of new ideas.

Diversity as Product Signal

Your Lived Experience Is a Product Design Superpower

"I recognized that almost everybody in tech assumed: one phone number + one device = one person. Growing up in Trinidad, I just knew that wasn't true."

Adriel was the first Black PM at Facebook — among ~30 total PMs at the time
His Trinidad background revealed a global blind spot in registration design: prepaid SIMs, shared devices, dual-SIM phones
His team: Black Trinidadian PM, Israeli female EM, Brazilian female tech lead, engineers from Russia, China, Slavic countries — global empathy baked in
Diverse teams catch product assumptions that homogeneous teams never question

The hiring case Diversity isn't just ethical — it's a product quality issue. Every assumption your team doesn't notice is a bug waiting to hurt users who don't fit the default mental model.

Contrarian

Algorithmic Product Myths — Debunked

✗ Feed all data to the algorithm and it'll do the right thing INSTEAD → ✓ Algorithms don't understand long-term effects, human responses, or your product intent. Techno-utopianism is a PM abdication, not a product strategy.

✗ Ops is an execution problem, not a product design problem INSTEAD → ✓ Operational control is a first-order design requirement. If your product can't be managed by operators in the field, the algorithm is ungovernable at scale.

✗ A diverse team is a nice-to-have for morale INSTEAD → ✓ Homogeneous teams ship products with global blind spots. Adriel's perspective on shared phones shaped Facebook's registration — and growth — for hundreds of millions of users.

✗ R&D labs just need space and resources to innovate INSTEAD → ✓ Innovation labs die from organizational organ rejection. The survival rule: make your work core to the mission, share the wins broadly, and never position yourselves as the only team that gets to invent.