Fake the AI Before You Build It
Never train a model for an MVP — prototype the AI's output and test demand first.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 92%
Nika's hard rule for early-stage teams: an MVP whose purpose is to prove market demand should contain zero real machine learning. Training burns weeks of expensive data-scientist and compute time to answer a question a Figma prototype answers in days. You fake the model's output, show it to users, and only invest in training once demand is proven and you have real (or adjacent) data to train on.
Origin
Marily Nika's advice to early-stage entrepreneurs who approach her asking how to train a model to validate a market — a Wizard-of-Oz style validation applied specifically to AI features.
Core principles
- 01An MVP exists to get buy-in and prove a market, not to prove technical feasibility.
- 02Training real models takes weeks of powerful machines and scarce data-scientist time.
- 03Only invest in a model where you already have data, or data from an adjacent product you can leverage.
How to run it
- 1
State what the AI would output
Write down exactly what the model would return to the user: the recommendation, the classification, the transcription, the score.
- 2
Build a Figma prototype that shows that output
Create a lightweight clickable prototype where the AI's output is hard-coded or hand-produced. Fake exactly what the AI is going to be doing.
Pro tip Hand-produce a handful of realistic outputs rather than perfect ones — you are testing whether users want the output, not how good it is.
- 3
Put it in front of users and read the reaction
Show the prototype to target users and stakeholders to see whether the smart feature is genuinely wanted and whether it changes behaviour.
- 4
Only then check your data position
If demand is proven, ask whether you already have data — or data from an adjacent product — that you can leverage. If not, you still should not train.
Watch out If the answer is 'we'd have to collect all the data first', the model investment is far larger than the validation you just did.
In the wild
Young early-stage entrepreneurs repeatedly ask Nika how to train a model to do X 'because we want to prove that there is a market'. Her answer is a flat no: use AI where you already have data or adjacent data — not for the MVP.
→ Founders redirect weeks of model-training effort into a fast prototype test that answers the market question directly.
Common mistakes
Using model training as market validation
Training answers 'can this be done?', not 'does anyone want it?'. Spending weeks of compute and data-science time to answer the wrong question is the most expensive way to learn nothing.
Burning your data scientist on a prototype
Data scientists who can train models on powerful machines are the scarcest resource on the team; spending them on an MVP that may be discarded destroys their leverage.
Is it for you?
Best for
Early-stage founders and zero-to-one PMs who believe their idea requires AI to be testable.
Not ideal for
Products where the model's accuracy IS the value proposition and users cannot judge the experience from a faked output.
From the transcript
“don't do it for your MVB it makes zero sense do not waste time of data scientists that can train models with us using powerful…”
“fake it create a little figma product type and just show it some users and just F what the AI is going to be doing”
From the episode
AI and product management
Marily Nika (Meta, Google)