## OCDevel AI Podcast — The ML Workflow and Train/Test Split The anchor episode of Phase 1: the end-to-end machine learning workflow as a loop, and the train/test split as the discipline that keeps your evaluation honest. Builds directly on the NumPy/pandas episode; unlocks every model that follows, because they all share the same fit/predict/score harness. ### Education segment (anchor) - **Workflow as a loop**, not a line: problem framing, data collection, EDA, cleaning, feature engineering, model choice, training, evaluation, tuning, deploy, monitor, iterate. Roughly 80% of project time is data + features, 20% modeling. Often drawn as [CRISP-DM](https://en.wikipedia.org/wiki/Cross-industry_standard_process_for_data_mining) with back-arrows. - **Generalization** is the goal: performance on unseen data from the same distribution. Training error is an optimistically biased estimate of true error, which is why you need held-out data. - **Overfitting vs underfitting**, capacity, and the bias-variance tradeoff (intuition only). - **The split**: 80/20, 70/30, 90/10 for big data; stratify for imbalance, never shuffle time series. Three-way split (train/validation/test) and k-fold cross-validation (Kohavi's 1995 case for 10-fold). - **Data leakage taxonomy**: preprocessing before split, target leakage (the Caruana et al. pneumonia case), temporal and group leakage, distribution shift. - **scikit-learn** facts verified against version 1.9.0: [train_test_split](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html), [cross_val_score](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.cross_val_score.html), the [cross-validation guide](https://scikit-learn.org/stable/modules/cross_validation.html), [common pitfalls](https://scikit-learn.org/stable/common_pitfalls.html), [DummyClassifier](https://scikit-learn.org/stable/modules/generated/sklearn.dummy.DummyClassifier.html), and the [Kaggle Data Leakage lesson](https://www.kaggle.com/code/dansbecker/data-leakage). - Worked examples: Iris with a dummy baseline (~0.33) vs logistic regression (~0.97); California housing with a mean baseline (R² ~0.0) vs linear regression (R² ~0.58); correct vs leaky scaling. ### News segment (week of June 18–22, 2026) - [Noam Shazeer leaves Google for OpenAI](https://www.cnbc.com/2026/06/18/google-gemini-co-lead-noam-shazeer-leaves-for-openai.html): Gemini co-lead and "Attention Is All You Need" co-author departs (also see [9to5Google](https://9to5google.com/2026/06/17/geminis-co-lead-is-leaving-google-to-join-openai/)). Google reportedly paid ~$2.7B in 2024 to bring him back. - [AI funding roundup](https://news.crunchbase.com/venture/biggest-funding-rounds-cybersecurity-defense-startup-ai-odyssey-leads/): Odyssey ($310M Series B, world models), Hydra Host ($100M), Bland AI ($50M), Radical Numerics ($50M). - Open-weights leads (reportedly): GLM-5.2, MiniMax M3, VibeThinker-3B, Nemotron 3 Ultra.