Phase 0, episode 3: the map of how machines learn. Episodes 1 and 2 covered what AI is and how it got here; this one subdivides machine learning by the form of its training signal into four paradigms. ## The four paradigms - **Supervised**: learn a function from inputs X to outputs y using labeled pairs a human supplied. Splits into regression (continuous y, like house price, loss = mean squared error) and classification (discrete y, like spam vs ham, loss = cross-entropy). The catch is labels are expensive; [ImageNet](https://www.image-net.org/) needed ~14 million images labeled via Amazon Mechanical Turk. Label quality caps model quality (the skin-cancer model that keyed on rulers). - **Unsupervised**: no labels, just structure. Clustering (k-means, customer segmentation), dimensionality reduction (PCA, t-SNE, UMAP), density estimation, anomaly detection. Pitfall: clustering always returns clusters, even from noise, so validate with the elbow method or silhouette score. - **Reinforcement**: an agent takes actions in an environment for a delayed scalar reward, learning a policy to maximize cumulative return. Core problems: exploration vs exploitation (multi-armed bandit, epsilon-greedy) and credit assignment (value functions, discounting by gamma). Examples: AlphaGo beating Lee Sedol 4-1 in 2016, Atari DQN, RLHF. Pitfall: reward hacking (the CoastRunners boat spinning in circles). - **Self-supervised**: labels manufactured from the data itself by hiding part and predicting it. Next-token prediction (GPT), masked language modeling (BERT), contrastive learning (SimCLR, MoCo, DINOv2). Yann LeCun calls it [the dark matter of intelligence](https://ai.meta.com/blog/self-supervised-learning-the-dark-matter-of-intelligence/) ([Lex Fridman #258](https://lexfridman.com/yann-lecun-3/)). SSL plus transformers plus scale produced [foundation models](https://arxiv.org/abs/2108.07258). Plus **semi-supervised** (a few labels, lots of unlabeled: FixMatch, MixMatch) and what an ML practitioner actually does day to day: framing, data work, baselines, evaluation, deployment. The "80% on data cleaning" claim is overstated; [Anaconda's surveys put data prep near 45%](https://www.bigdatawire.com/2020/07/06/data-prep-still-dominates-data-scientists-time-survey-finds/) ([Leigh Dodds breaks down the myth](https://blog.ldodds.com/2020/01/31/do-data-scientists-spend-80-of-their-time-cleaning-data-turns-out-no/)). ## News (June 8-11, 2026) Apple's WWDC keynote unveiled a Gemini-powered Siri and an Extensions system letting users pick ChatGPT, Gemini, or Claude. Anthropic released Claude Fable 5 at roughly half the prior price with a reroute-to-Opus safety mechanism. OpenAI shipped its "Dreaming" memory upgrade and announced Oracle cloud access. An ICLR 2026 paper compresses models during training using Hankel singular values.