OCDevel AI Podcast

Linear Regression: Your First Real Model, From a scikit-learn One-Liner to the Intuition Underneath

OCDevel AI Podcast — Linear Regression (Phase 1: Your First Real Model)

This episode opens with a quick news brief, then settles into the anchor tutorial: linear regression, the first real model in the course. Intuition first, then a hands-on scikit-learn workflow.

News brief (June 22–28, 2026)

  • OpenAI GPT-5.6 "Sol / Terra / Luna" — limited release June 26. A three-tier family (flagship / balanced / fast-cheap) with new "max" and "ultra" reasoning modes; reported pricing and gains. 9to5Mac
  • Anthropic Claude Tag — June 23. A "virtual teammate" you @-mention in Slack; runs on Opus 4.8. Anthropic · Fortune
  • SpaceX × Reflection AI — June 22. ~$6.3B compute deal for Nvidia GB300s at Colossus 2 near Memphis. TechCrunch · CNBC
  • OpenAI "Daybreak" / GPT-5.5-Cyber — June 22. Defensive-security model, Codex Security plugin, "Patch the Planet." Axios · Infosecurity
  • DeepMind talent moves — Jumper to Anthropic, Shazeer to OpenAI. TechCrunch · Fortune
  • ByteDance Seedance 2.5 — June 23, reportedly 30-second single-pass video.

Tutorial: Linear Regression

  • Regression vs classification; simple vs multiple regression; the hypothesis y = w·x + b and why that dot product is the atom of everything downstream.
  • The scikit-learn one-liner: fit / predict / score on LinearRegression (stable 1.9.0). Worked on fetch_california_housing (not the removed Boston set), honest test R² ≈ 0.6.
  • Reading coefficients (ceteris paribus), the units/scaling caveat, and why magnitude ≠ importance.
  • What fitting does: residuals, OLS, MSE, and why squared (penalty, differentiability, convexity, the Gaussian/MLE link). Closed-form normal equation vs gradient descent.
  • Evaluation: R², MSE, RMSE, MAE — computed on held-out test data.
  • Assumptions (the classic five) and the master diagnostic, the residual plot.
  • Pitfalls: correlation ≠ causation, extrapolation, outliers, multicollinearity, and Anscombe's quartet — always plot your data.
  • History: least squares (Legendre 1805, Gauss 1809) and Galton's "regression toward the mean."
  • Forward: logistic regression → the single neuron → neural networks.