The Model Is A Small Part
Across real ML teams, the actual model choice often accounts for a small portion of the work. What consumes time is:
- Understanding the problem and the decision the model will inform.
- Sourcing, cleaning and labelling data.
- Defining evaluation metrics that reflect real outcomes.
- Deploying, monitoring and iterating in production.
The Lifecycle
A useful mental model of the ML lifecycle looks like this:
- Problem framing — what decision are we improving?
- Data collection — do we have the right data, at the right resolution, ethically sourced?
- Feature preparation — how do we represent the data so a model can learn?
- Modelling — start simple, add complexity only when justified.
- Evaluation — offline metrics and business-level checks.
- Deployment — serving, latency, cost.
- Monitoring — data drift, performance decay, edge cases.
Why "Simple First" Wins
Strong ML teams almost always begin with the simplest baseline that could work. This gives them:
- A reliable point of comparison.
- Faster iteration.
- A clearer view of what the problem actually is.
Skills That Actually Matter
Beyond mathematics and modelling, engineers who succeed in ML develop:
- Strong data intuition.
- Careful evaluation habits.
- Software engineering discipline.
- The ability to communicate uncertainty.
These are exactly the skills that survive shifting model architectures.
