Latent Demand
Latent demand is the product signal embedded in how users already behave — often by hacking or misusing an existing product to meet a need that product was never designed for. Boris Cherny calls it ‘the single most important principle in product.’ He identifies a second, modern form applicable specifically to AI products.
Form 1 — User latent demand (traditional)
Look at what users are already doing, even when it requires going out of their way. The signal is in the workaround, not the feature request.
Examples:
- Facebook Marketplace — before it existed, 40% of Facebook Groups posts were people buying and selling. The team built Marketplace by formalising what users were already doing.
- Facebook Dating — 60% of profile views on Facebook were non-friends of the opposite gender, indicating existing dating-like behaviour on a platform not designed for it.
- Claude Code → Cowork — non-engineers (data scientists, analysts) were learning to open a terminal just to use Claude Code for SQL queries, genomics, MRI analysis, growing tomatoes. A data scientist (Brendan) appeared one day using Claude Code in a terminal; within a week, the whole data science team was doing the same. Anthropic built Cowork — a desktop app without the terminal — to serve this latent demand directly.
The practical rule: when users jump through hoops to use your product for a purpose it wasn’t designed for, that is extremely strong evidence that a purpose-built product will succeed.
Form 2 — Model latent demand (modern)
A newer framing specific to AI products: look at what the model is trying to do, not just what users are doing.
The standard pattern for building with LLMs is to treat the model as a component: ‘here is my system; model, you perform step 3.’ This over-constrains the model.
Boris’s inversion: make the product be the model. Give it tools. Give it a goal. Let it determine the sequence of tool calls. In ML research, building in this way is described as being ‘on distribution’ — not fighting the model’s natural behaviour.
‘Latent demand of what the model wanted to do. We wanted to expose it. We wanted to put the minimal scaffolding around it.’
This is directly connected to Bitter Lesson: scaffolding that constrains the model’s planning yields at most 10–20% improvement over letting it plan freely, and these gains are erased by the next model.
Relationship to other concepts
Latent demand (form 2) is the product-layer analogue of the Bitter Lesson. Both argue that you get better results by exposing general capability than by constraining it. Latent demand (form 1) is a related idea to Product Taste — the ability to distinguish a real user need from an articulated preference — but grounded in observed behaviour rather than user empathy.
Nikita Bier: organic discovery as the strongest signal
Nikita Bier on Viral Consumer Apps, Latent Demand, and Building for Teens provides the most vivid illustration of Form 1 in the consumer social context.
For tbh, the app appeared at #1 in the US App Store in Arabic before the team understood what had happened. Users in Arabic-speaking communities had found and propagated it through channels the team had not designed. Bier’s framing: this is the strongest possible proof of latent demand — people discovering a product before the product finds them.
The operational test he derives: if your product is spreading through channels you did not create, in demographics you did not target, you are sitting on latent demand. If every new user arrived through a paid or deliberate channel, the demand is manufactured rather than latent.
This extends Form 1 from the ‘user workaround’ signal (users hacking an existing product) to the ‘organic discovery’ signal (users finding a new product in unexpected ways). Both are evidence of a pre-existing desire looking for satisfaction.