Machine Learning

Machine learning (ML) is the field of computer science focused on building systems that learn patterns from data rather than following explicitly programmed rules. It encompasses a broad spectrum from classical statistical methods (linear regression, decision trees, clustering, random forests) through deep learning (neural networks with many layers) to the modern generative AI and foundation models that dominate current attention. ML is the parent field within which generative AI is a recent and currently dominant subcategory. Most ML work today is still classical or traditional deep learning; generative AI is the most visible but not the totality of ML.

The main categories of machine learning:

Supervised learning: train on labeled examples (input → output pairs). Most common.

• Examples: image classification, spam detection, credit scoring, churn prediction.
• Algorithms: linear/logistic regression, decision trees, random forests, gradient boosting (XGBoost), neural networks.

Unsupervised learning: find patterns in unlabeled data.

• Examples: customer segmentation, anomaly detection, recommendation systems.
• Algorithms: k-means clustering, PCA, autoencoders.

Reinforcement learning: learn through trial and error with reward signals.

• Examples: game-playing AI (AlphaGo), robotics, RLHF training of LLMs.
• Algorithms: Q-learning, policy gradients, PPO.

Semi-supervised: combine labeled and unlabeled data.

Self-supervised: generate labels from the data structure itself (modern LLM pre-training).

The machine learning stack:

Data layer: collection, labeling, storage, versioning, cleaning.

Feature engineering: transforming raw data into model-ready features.

Model training: experiment tracking, hyperparameter tuning, distributed training.

Model deployment: serving, A/B testing, monitoring.

Model operations (MLOps): managing models in production, retraining, drift detection.

Classical ML vs deep learning vs generative AI:

Era	Approach	Examples
1950s-2000s	Statistical ML	Linear regression, SVMs, decision trees
2010s	Deep learning	Image recognition (CNNs), translation (RNNs/LSTMs)
2017+	Transformer architecture	BERT, GPT
2020+	Foundation models	GPT-3, Llama
2022+	Generative AI mainstream	ChatGPT, DALL-E, Midjourney

Each era didn't replace the previous; they coexist. Classical ML still dominates for tabular data (financial models, credit scoring, predictive maintenance). Deep learning dominates for unstructured data (images, text, audio). Foundation models dominate for general-purpose tasks.

Where ML matters for startups beyond LLMs:

Forecasting: time-series prediction (revenue, demand, churn).

Recommendation systems: classical collaborative filtering still drives most recommendations.

Anomaly detection: fraud, security, operations monitoring.

Image processing: computer vision tasks (manufacturing QA, medical imaging).

Optimization: routing, scheduling, resource allocation.

Predictive maintenance: when does equipment fail?

Personalization: tailoring content/experience per user.

The startup ML talent question:

Classical ML engineer: $150K-$300K base + equity. Builds models for forecasting, recommendation, etc.

ML engineer (deep learning): $200K-$400K base + equity. Trains/serves models, builds ML systems.

ML research scientist: $250K-$500K+ base + significant equity. Top-tier research talent.

Applied AI engineer: $200K-$400K. Integrates foundation models into products, fine-tuning, prompting.

The right talent depends on what you're building. Most startups don't need research scientists; they need applied AI engineers who can integrate foundation models effectively.

The fundamental ML stack today:

Foundation model APIs (OpenAI, Anthropic, Google): for generative AI work.

Open-source foundation models (Llama, Mistral): for self-hosted needs.

Classical ML frameworks (scikit-learn, XGBoost): for tabular data.

Deep learning frameworks (PyTorch, TensorFlow): for custom model training.

MLOps platforms (MLflow, Weights & Biases, Comet): for experiment tracking and deployment.

What founders get wrong: Conflating "machine learning" with "generative AI" and assuming all ML problems should be solved with LLMs. The right discipline: understand the breadth of ML; use the right tool for the problem; classical ML is often the right answer for structured data and forecasting.

Related: Generative AI · Foundation Model · Large Language Model · Training Data · ML Engineer

FAQ

What is machine learning?
The field of computer science focused on building systems that learn patterns from data rather than following explicitly programmed rules. Encompasses classical ML (regression, decision trees, clustering), deep learning (neural networks), and generative AI (LLMs, image models). Generative AI is a subcategory, not the whole field.

What's the difference between machine learning and generative AI?
Machine learning is the broad parent field. Generative AI is a recent subcategory focused on creating new content. ML also includes classical methods (forecasting, recommendation, anomaly detection) that don't generate content. Most production ML at companies is classical, not generative.

What are the main categories of machine learning?
Supervised learning (labeled training data), unsupervised learning (patterns in unlabeled data), reinforcement learning (trial-and-error with rewards), semi-supervised (mix of labeled and unlabeled), self-supervised (labels from data structure). Each category fits different problems.

Should I use LLMs or classical ML for my problem?
Match tool to problem. LLMs: text/image/code generation, conversation, general reasoning, unstructured data. Classical ML: forecasting, recommendation, anomaly detection, structured/tabular data. Don't use LLMs for everything; classical ML often produces better results for structured data.