Why Learning Machine Learning Still Matters in the Age of Giant AI Models

1. The Power of Large Models Relies on Machine Learning

Large language models are impressive and becoming stronger, thanks to advances in machine learning such as Transformer‑based deep neural networks, massive unsupervised pre‑training, instruction tuning, RLHF, and multimodal fusion.

Today’s models are the culmination of two decades of machine‑learning research, not a replacement for it.

Because large models depend on machine‑learning breakthroughs, we must not invert causality: machine learning enables large models, not the other way around.

Is it still necessary to learn "machine learning"?

2. Large Models Are General, Machine Learning Is Custom

Large models require massive data, and in highly specialized domains high‑quality training corpora may be scarce.

When data are limited or highly structured, traditional machine‑learning methods (e.g., decision trees, SVM, random forests) and careful feature engineering still provide efficient and stable advantages.

In industrial or research settings that demand interpretability, robustness, or resource constraints, lightweight traditional methods become indispensable, highlighting the importance of machine‑learning’s "custom" capability.

Understanding machine‑learning fundamentals also lets us choose or develop new models instead of blindly delegating everything to a massive, compute‑hungry model.

3. Without Machine‑Learning Knowledge You Can’t Use Large Models Effectively

Large models are powerful but not a universal solution; lacking knowledge of underlying principles, data processing, and evaluation can turn them into opaque black boxes that mislead users.

For structured data or specific industrial scenarios, prompt engineering or domain‑specific post‑processing is required, and without grasping data features and algorithm mechanics you cannot decide when to apply a large model directly or combine it with traditional algorithms.

Fine‑tuning large models or coordinating them with smaller ones also demands solid machine‑learning foundations: selecting appropriate data, avoiding over‑fitting, ensuring training stability, and evaluating performance iteratively.

Thus, machine learning helps us understand both the capabilities and limitations of large models, enabling us to extract their maximum value.

4. Machine Learning Remains the Best Training Ground for "Math + Programming + Thinking"

Learning and practicing machine learning is essentially a comprehensive exercise in mathematics, programming, and problem‑analysis thinking. It requires solid knowledge of linear algebra, probability, statistics, as well as hands‑on coding for data preprocessing, model building, and hyper‑parameter tuning, and the ability to abstract and reason about complex problems.

Mathematics: Algorithms and models depend on rigorous mathematical theory for interpretability and stability; understanding gradient descent or matrix solutions is essential.

Programming: Every stage of machine‑learning pipelines—from data handling to model deployment—needs code; even when large models assist development, you must verify, debug, and improve generated code.

Data Thinking: The core skill is to transform raw data or vague requirements into analyzable, solvable models, structuring chaos, quantifying intuition, and designing systematic experiments—abilities that large models cannot provide directly.

While some low‑level implementations can be automated, machine learning remains the optimal arena to solidify these foundational abilities for the era of large models.

5. Course Design: Trim the Trivial, Keep the Core

When planning curriculum, I prefer to "reduce tedious technical details while preserving key principles and thinking training".

Traditional courses often spend extensive time on manual implementation of decision trees or derivations of regression algorithms; in the age of large models these steps can be automated, but we must still emphasize why an algorithm is chosen, its applicability boundaries, data handling, evaluation, interpretation, and pitfalls.

Only by mastering these core concepts will students remain confident and adaptable when newer, more powerful models appear, allowing them to select, adapt, or create tools without fearing replacement.