LLenny's Podcast
← All frameworks
StrategyMarily Nika (Meta, Google)

The Model Launch Bar

With probabilistic products, the PM — not the scientist — decides what accuracy is good enough to ship.

Difficulty
Moderate
Time to result
~months to results
Steps
4
Confidence
85%

AI outputs are probabilities, not answers: a model says 'that's a rhino, and I'm 70% sure'. That means shipping is no longer a binary works/doesn't-work call — someone has to decide where the accuracy bar sits for real users. Nika insists that decision is the PM's, and pairs it with a second PM-owned decision: whether to use an off-the-shelf model or invest in proprietary data, because everyone buying the same ready-made dataset ships the same quality.

Origin

Marily Nika's framing of what makes the AI PM role distinct, from shipping ML products at Google and Meta.

Core principles

  • 01A model takes an input and outputs a prediction plus a confidence — never a certainty.
  • 02Where the quality bar sits is a product decision about user tolerance, not a technical one.
  • 03If everyone trains on the same purchasable dataset, everyone's quality is identical — differentiation comes from your own data.
  • 04The PM must also own the monetisation bridge from research output to a paid use case.

How to run it

  1. 1

    Define the user-facing failure mode

    Work out what happens to the user when the model is wrong. A wrong movie recommendation and a wrong medical flag have wildly different tolerances.

  2. 2

    Set the accuracy bar explicitly, and own it

    Decide the number: is 70% accurate good enough? 80%? State the bar before training finishes so the team isn't negotiating it against sunk cost.

    Pro tip Nika treats this as the fun part of the role — it's a decision nobody has tackled before, and it's entirely yours.

    Watch out Do not delegate this to the data scientist; they optimise the metric, they don't own the user's tolerance for being wrong.

  3. 3

    Decide off-the-shelf vs. proprietary data

    If you're a large company offering a service where quality is the differentiator, train on more and more diverse data of your own. Buying a ready-made dataset from an agency means shipping the same quality as everyone else who bought it.

    Watch out Agencies sell the same data packages to your competitors — parity, not advantage.

  4. 4

    Bridge research to a monetisable use case

    Take the research scientist's amazing idea and figure out how it actually monetises — willingness-to-pay research (what's the minimum you'd pay, the maximum, what would you want if you paid) is the PM's job, not the scientist's.

In the wild

The cat-or-dog launch decision

Nika's recurring illustration: the classifier says it's 70% sure a photo is a cat. Is 70% good enough to put in front of users, or is the bar 80%? Nobody but the PM can answer that.

The accuracy threshold becomes an explicit, defensible product decision rather than a number the data scientist happened to land on.

ChatGPT's willingness-to-pay survey

Nika points to the sign-up form circulating in early 2023 asking users what the minimum they'd pay would be, what the maximum, and what they'd want to see if they paid — an example of a PM bridging free research capability to a priced product.

A paid tier launched shortly after, priced from that demand signal rather than from cost.

Common mistakes

Treating model accuracy as an engineering metric

Accuracy targets set without a view on user tolerance either block a shippable product forever or ship something users don't trust. The bar is a product judgment.

Buying the same dataset as your competitors and expecting an edge

If everyone takes the exact same ready-made dataset, every company's model quality converges. Diversifying and collecting your own data is the only route to a quality advantage.

Is it for you?

Best for

PMs shipping their first probabilistic feature and deciding when it's good enough for users.

Not ideal for

Safety-critical domains where the acceptable error rate is set by regulation, not by product judgment.

From the transcript

it's totally your responsibility as a p to decide okay the recognition of whether this phold is a cat or a doob is good enough…

18:00

there are agencies that are selling data packages of data that are ready so you can get them and train your models but the question…

17:30

if you're a PM you need come up with ways to actually

34:00

From the episode

AI and product management

Marily Nika (Meta, Google)