What Kind of Product Manager Drives Algorithm Engineers Crazy?

1. Algorithm engineers hate more than "you don’t understand tech"

Many non‑technical PMs don’t lack coding ability; they disregard the underlying logic of algorithms, demanding outcomes like “more precise recommendations” without checking data availability, feature feasibility, or model viability. This input‑output‑only communication wastes resources and erodes trust.

2. Ignoring this makes you "dangerous" in the team

1. Trust collapses

When algorithms are repeatedly given unrealistic requests, they acquire a "unreliable" label and switch to defensive mode. A two‑week task can be stretched to a month because data is insufficient, and teammates begin to assume every request has hidden pitfalls.

2. Produce useless output

Teams sometimes chase marginal accuracy improvements that do not affect business conversion. After weeks of effort, the product’s KPI remains unchanged, and algorithms can only say, "the requirement was yours, we just implemented it." Such projects quickly diminish a PM’s influence.

3. Avoid these two traps: effort invites trouble

Trap 1: Learning algorithm buzzwords without depth

Memorizing terms like loss function, gradient descent, or over‑fitting does not make a PM a competent guide. Acting as a half‑baked theorist leads to misguided directions, similar to a layperson telling a chef how to cook.

Trap 2: Over‑humble, no stance

PMs who constantly ask, "Is this okay?" or "Whatever you say" provide no clear direction. Algorithms find it hardest when the PM cannot even articulate what they want, leading to frustration and reduced collaboration.

4. Three practical methods to become a "great teammate"

1. Drop the black‑box mindset, talk with "features"

Instead of saying, "I need a recommendation system," start by listing:

What raw data do we have?

Which features can we extract?

Which features are truly useful for the target?

Real example: when improving car‑pool order‑success rate, we did not ask for a prediction model directly. We first enumerated user ride frequency, order time, distance, weather, and traffic conditions. Discussing these features made the algorithm team’s eyes light up and aligned expectations instantly.

2. Reject vague slogans, quantify metrics

Algorithms dislike vague goals like “better accuracy” or “more beautiful output.” Replace them with concrete numbers, e.g.:

Increase Top‑3 recommendation click‑through rate by X %.

Boost GMV without decreasing retention.

Achieve style‑consistency score ≥ Y and generation latency ≤ 5 s.

Clear, measurable targets give algorithms a precise target instead of guessing the PM’s intent.

3. Respect experiment cycles, give algorithms failure space

Algorithm work is a scientific experiment, not a one‑day page build. Schedule must include:

Offline experiment time.

Model tuning time.

Online A/B testing (gray‑release) time.

Ask the algorithm team, "What is the baseline? How many experiment rounds are needed to see effect?" and, in front of leadership, help them “record the timeline.” Understanding their workload prevents unnecessary pressure and encourages them to stay up late to support you when issues arise.

Conclusion: Algorithms are cold, collaboration is warm

Product work can become a mechanical task of chasing metrics, but every line of code and data point represents real people. Before raising a requirement, ask if the data volume is sufficient; after launch, thank the algorithm team with a coffee or a public acknowledgment. Simple gestures turn a cold algorithm partnership into a warm, effective collaboration.