Machine learning

What is Machine learning?

Traditional software is a list of rules a person wrote down. Machine learning flips that: you show a system many examples of inputs paired with the right answers, and it works out the rules itself. Nobody codes "this email is spam." You feed in a few hundred thousand emails already labeled spam or not spam, and the model learns which patterns tend to go with which label.

The output of that process is a model, which is really just a large set of numbers that turns an input into a prediction. Training is expensive and slow. Using the trained model, called inference, is usually fast and cheap. That split matters for founders, because it means you can often rent someone else's training and pay only for inference.

What makes machine learning different from ordinary code is that it is probabilistic. It does not return "correct," it returns "most likely." A recommendation engine that is right 70 percent of the time can be enormously valuable, while a billing system that is right 70 percent of the time is a lawsuit. Knowing which side of that line your feature sits on is the whole skill.

Types of machine learning

Type	What it learns from	Typical startup use
Supervised	Examples with known correct answers	Spam filtering, lead scoring, churn prediction
Unsupervised	Unlabeled data, finding structure	Grouping users, anomaly detection
Reinforcement	Trial, error, and a reward signal	Pricing, routing, game agents

Most products that describe themselves as AI today sit on top of large pretrained models reached through an API rather than anything trained in house. That is a legitimate and often correct choice for a small team.

Why machine learning matters for startups

The honest founder question is not "should we use machine learning" but "does this problem have a shape machine learning fits." It fits when the rules are too messy or too numerous to write by hand, when you have enough historical examples, and when being occasionally wrong is acceptable.

It also changes what a competitive moat looks like. Anyone can call the same model API you call. What they cannot easily copy is your data: the labeled examples your users generate through normal use. A team of three with a narrow dataset nobody else has can beat a much larger team using only generic models.

The trap is cost structure. Inference is a per-request expense that scales with usage, which makes it behave more like a hardware bill than like software. Model that into your pricing before launch, not after.

Machine learning in practice

Say you run a small support inbox tool and want to auto-route tickets to the right team. The tempting move is to train a classifier immediately. The cheaper move is to write twenty keyword rules first, measure how often they are right, and use the corrections your customers make as labels.

After four months you have around 40,000 tickets with a human-confirmed team attached. Now a simple classifier has something real to learn from, and you can ship it as a suggestion that agents accept or reject with one click. Every rejection becomes another training example. The rules version got you to market in a week, and it quietly built the dataset that made the model possible.

Common mistakes

• Starting with the model instead of the problem. If a lookup table or a few if statements solve 80 percent of the case, ship that first and learn what the remaining 20 percent actually looks like.
• Ignoring the data collection loop. A model with no mechanism for capturing corrections stops improving the day it launches. Design the feedback path before the model.
• Testing on the data you trained on. Scores look wonderful and mean nothing. Hold out a slice of data the model has never seen.
• Hiding uncertainty from users. Presenting a guess as a fact destroys trust the first time it is wrong. Show confidence, and let people override.
• Forgetting that inference costs money per call. A free tier with unlimited model calls can turn a good week of growth into a painful invoice.

Related concepts

Machine learning is one branch of artificial intelligence, and it depends heavily on having enough clean examples, which is where big data enters the picture. For a small team, the practical route is to wrap a model in a narrow minimum viable product, measure whether the predictions change user behavior with disciplined A/B testing, and only then invest in training anything custom.