The ML Landscape
Context
Section titled “Context”Your engineering team proposes building a recommendation feature with “unsupervised learning.” In the next sentence, someone drops “reinforcement learning.” The Slack thread now mentions “self-supervised pre-training.”
You don’t need to train a model. But you need to understand which type of ML solves which problem — otherwise you can’t evaluate the proposal or ask the right question.
Concept
Section titled “Concept”Three Programming Paradigms
Section titled “Three Programming Paradigms”Andrej Karpathy distinguishes Software 1.0 from 2.0. For PMs, this is the most important foundational distinction:
- Traditional programming (Software 1.0): A human writes rules.
if order_status == "shipped": show_tracking() - Machine Learning: The machine learns rules from labeled examples. You provide 10,000 categorized support tickets, the model learns the mapping.
- Deep Learning: The machine learns directly from raw data — no manual feature engineering. Tesla Autopilot processes camera images; no human defines “what a traffic light looks like.”
Neural Networks — PM Level
Section titled “Neural Networks — PM Level”A neural network has input layers, hidden layers, and output layers. “Deep” means many hidden layers. Each connection has a weight learned during training. Large frontier models have hundreds of billions to over a trillion parameters — exact numbers are rarely officially confirmed (OpenAI, Google, and Anthropic don’t publish parameter counts).
You don’t need to know the math. But you need to know: more parameters = more capacity, but also more cost and more data required to train.
Five Types of Machine Learning
Section titled “Five Types of Machine Learning”| Type | How it learns | PM example |
|---|---|---|
| Supervised | Labeled data (input → known output) | Spam detection, churn prediction |
| Unsupervised | Patterns in unlabeled data | Customer segmentation, anomaly detection |
| Semi-Supervised | Small labeled + large unlabeled set | When labeling is expensive but raw data is abundant |
| Self-Supervised | Model generates its own labels from data | LLM pre-training (predict the next token) |
| Reinforcement Learning | Trial-and-error with rewards | RLHF for chatbots, process optimization |
Classification vs. Generation
Section titled “Classification vs. Generation”- Discriminative models learn boundaries: Is this email spam or not? They compute P(label|data).
- Generative models learn the data distribution: What does a typical email look like? They compute P(data).
Your PM rule of thumb: Sort/label → discriminative. Create/produce → generative. Both → hybrid pipeline. Gmail Smart Reply does exactly this: classify intent (discriminative), then generate a reply (generative).
How LLMs Develop “Taste”
Section titled “How LLMs Develop “Taste””“Pre-training = reading every cookbook. Fine-tuning = watching a chef. RLHF = developing taste.”
- RLHF (Reinforcement Learning from Human Feedback): Pre-train → build a reward model from human feedback → fine-tune with PPO or DPO (Direct Preference Optimization — equally common since 2024).
- Constitutional AI (Anthropic): The model critiques itself against defined principles — scalable without thousands of human raters.
Framework
Section titled “Framework”The Data-Task Matrix — your decision grid for ML approaches:
| Labeled data available | Unlabeled data only | |
|---|---|---|
| Classify/Predict | Supervised Learning | Unsupervised or Semi-Supervised |
| Generate/Create | Fine-tuned Generative Model | Foundation Model + Prompting/RAG |
| Optimize a process | Reinforcement Learning | Reinforcement Learning |
How to use the matrix: First determine your task (row), then your data situation (column). The cell tells you where to start.
Scenario
Section titled “Scenario”You’re a PM at a streaming service. The goal: better recommendations. Your data team presents three proposals:
- Option A: Supervised model trained on historical ratings (4.2M ratings available)
- Option B: Unsupervised clustering of users by viewing behavior, then rule-based recommendations
- Option C: Foundation model (LLM) with prompting: “Based on these movies, recommend similar ones”
Context:
- Netflix uses a hybrid: collaborative filtering (unsupervised) + supervised prediction
- Your budget doesn’t allow fine-tuning a foundation model
- You have 4.2M labeled ratings + viewing histories from 800,000 users
Decide
Section titled “Decide”Which approach would you choose?
The best decision: Combine Option A + B — a hybrid approach, following the Netflix playbook.
Why:
- You have labeled data (ratings) → supervised works. The Data-Task Matrix is clear: classify + labeled data = supervised.
- Unsupervised clustering surfaces additional signals (users with similar viewing behavior) that improve the supervised model.
- Option C (LLM with prompting) sounds modern, but it has no personalization based on your user data — and costs significantly more per request than a specialized recommendation model.
What many get wrong: Reaching for an LLM for everything, when a specialized ML model is cheaper, faster, and better suited for the use case.
Reflect
Section titled “Reflect”- ML is not a monolith. Supervised, unsupervised, self-supervised, and RL solve fundamentally different problems. The wrong choice costs months.
- Your data dictates the approach. Don’t pick the technology and then look for data — let your available data determine what’s possible.
- Generative is not always better. Discriminative models are often cheaper, faster, and more reliable for classification tasks. Not every problem needs an LLM.
- Hybrid is often the answer. The best production systems combine multiple ML types in a single pipeline.
Sources: Andrej Karpathy “Software 2.0” (2017), Chip Huyen “Designing Machine Learning Systems” (O’Reilly, 2022), Anthropic “Constitutional AI” (2023), OpenAI “Training language models to follow instructions with human feedback” (2022)