Hi Reader,

A reader asked me the following question:

Excellent question! As someone who just published a book on Machine Learning, I clearly believe there is still great value in learning how to do supervised Machine Learning yourself rather than just feeding the problem and dataset to an LLM. Here's why:

1. LLMs perform well on software engineering tasks (for example) because there is wealth of high-quality training data, correctness can be verified by deterministic automated tests, and there's a tight executable feedback loop (i.e., the code either runs or it crashes). With Machine Learning, on the other hand, it's easy to make subtle data-related or methodological mistakes that won't cause the program to crash, and thus it's essential for a knowledgeable human to be involved in order to catch those mistakes.
2. Unless you're using a self-hosted LLM or an API with strict privacy controls, transmitting your data to an LLM means giving potentially sensitive information to a third party whose data handling practices may not meet your requirements. Classical Machine Learning models are easily self-hosted, which means that you remain in complete control of your data.
3. LLMs are expensive and relatively slow, which makes them ill-suited to environments in which predictions need to be made in near-real-time or at very high volumes. In contrast, some classical Machine Learning models (trained and optimized by experienced humans) can make predictions very quickly and at extremely low cost, making them well-suited to those environments.
4. Some predictive models need to be deployed on devices that have limited memory, power constraints, or are not connected to the Internet. In those cases, LLMs are not feasible, whereas some classical Machine Learning models can be pruned or compacted to meet those constraints.
5. Although LLMs have made great strides in terms of their mathematical fluency through the use of tools, code interpreters, and chain-of-thought reasoning, they still require additional verification layers to ensure the correctness of their outputs. For any high-stakes application that uses Machine Learning, human guidance is still recommended in order to ensure trustworthy results.
6. There is much more to the Machine Learning lifecycle than just the initial model training, such as monitoring for data drift, handling schema changes in incoming data, running A/B tests, and scheduling retraining. These steps cannot be autonomously managed by LLMs and still require human expertise.
7. When LLMs are used to directly make predictions (rather than writing code to be separately executed), they effectively function as "black boxes" with no built-in explainability. In contrast, some classical Machine Learning models can be inspected in order to understand their output, which is especially important in regulated industries in which model audits may be required.
8. Although LLMs excel with unstructured data, the vast majority of business intelligence is stored in structured, tabular datasets. Classical Machine Learning approaches thrive with tabular data, and will often outperform LLMs in terms of accuracy and efficiency.

You may have heard of Tabular Foundation Models (TFMs), which is an emerging new approach to Machine Learning that has some similarities to LLMs. While TFMs automate parts of the Machine Learning workflow, using them still requires a human to understand problem framing, data leakage, evaluation metrics, and many other Machine Learning fundamentals.

For all of the reasons above, I strongly believe that a foundational understanding of supervised Machine Learning will remain valuable for the foreseeable future.

Do you agree or disagree with my assessment? Are there any key points that I missed? Reply and let me know!

- Kevin