Marily Nika on AI Product Management, Building with ML, and the Future of PM

Marily Nika on AI Product Management, Building with ML, and the Future of PM

transcriptlennymarily-nikaaimachine-learningproduct-managementmeta

Marily Nika on AI Product Management, Building with ML, and the Future of PM

Marily Nika — Lenny’s Podcast · ~early 2023 · Source

Marily Nika, product lead at Meta (Metaverse, avatars, and identity), previously Google (Google Glass, computer vision, speech recognition), and creator of the most-subscribed AI and PM course on Maven, explains what product managers need to understand about machine learning to be effective — without needing to code. Recorded at a moment when ChatGPT had just launched a paid tier, the episode is an early-wave introduction to AI product thinking.

Key ideas

  • The shiny object trap. The most common AI mistake is building AI because everyone is building AI. The diagnostic: is there a specific problem with a specific pain point that a smart automated system would solve better than any current alternative? Without that anchor, AI investment becomes expensive exploration. The generalist PM ships the right product; the AI PM solves the right problem — the emphasis shifts from output to problem definition.

  • Don’t use AI for your MVP. Training a model takes weeks of compute time and expert attention; using one for a proof of concept is wasteful. Fake what the AI would do with a Figma prototype, validate that users respond to the capability, then invest in building the real model. A majority of AI investments fail — this sequencing reduces the cost of that failure to near zero.

  • Every PM will become an AI PM. Not because they will train models themselves, but because every product will embed personalisation, recommendation, and automation — all of which require working alongside research scientists. The skill the role demands is comfort with uncertainty: ML projects do not have sprint-cycle certainty; a model can train for months and not answer the hypothesis it was designed to test. PMs who need clear launch dates and defined deliverables will struggle.

  • When to build vs buy a model. If all companies train on the same publicly available dataset, their models converge on identical quality. Building your own model is warranted when proprietary data volume exists and product quality is the differentiator — speech recognition, large-language products, recommendation systems at scale. For most applications, starting with an API (GPT, AutoML, Google’s pre-trained models) and layering proprietary data on top is faster and cheaper.

  • Learning to code changes your thinking. Understanding how a tool was built gives a PM the confidence to push on its limits and make intelligent trade-offs. The piano analogy: learning classical scales before the songs you love feels wrong, but it informs everything. Nika recommends non-programmer PMs take at least one coding course to acquire the mental model, not to become engineers.