AI Writes My Code Now. So What Am I For? | An ML Engineer’s Honest Take

Summary

A machine learning engineer reflects on feeling less valuable as AI increasingly writes code. The work has shifted: AI often creates the first draft, while engineers review, test, and merge it. Although debugging and coding can be frustrating, solving bugs by hand used to provide a sense of accomplishment—one that can feel diminished when AI does much of the implementation.

The speaker sees conflicting expectations in the industry. Companies encourage employees to use tools such as ChatGPT, Microsoft Copilot, and Claude, while technical interviews may still require coding by hand. Machine learning roles are also expanding beyond traditional model-building and tools such as scikit-learn and PyTorch, with new expectations around AI agents and running large language models locally.

The speaker argues that engineers remain valuable not because they can outpace AI, but because they bring three capabilities:

1. Communication: Engineers must explain the problem, justify an AI-generated solution, and communicate its trade-offs. For example, adding a cache to speed up an API may risk serving outdated data. “The AI did it” is not an adequate explanation.
2. Responsibility and ownership: AI can complete many tasks quickly, but engineers remain accountable for unintended effects, such as a change to a shared function breaking another part of an application. The speaker recommends limiting AI work to what they can meaningfully understand and review.
3. Curiosity and critical thinking: Fundamentals still matter for evaluating AI’s output. In a fraud-detection example, a model could achieve 99% accuracy simply by labeling every transaction “not fraud” if only 1% of transactions are fraudulent. Understanding class imbalance helps an engineer recognize that accuracy alone is misleading and ask how many fraud cases the model actually detects.

The speaker’s overall view is that AI should be used, but at a pace that lets engineers remain informed and accountable. They acknowledge that anxiety about changing roles is real, while suggesting that learning fundamentals and questioning AI’s choices can make engineers more effective.

Guides, Reviews, and Examples

• Practical guidance: Use AI only for work you can understand and review; question its reasoning and be prepared to explain the result.
• Machine learning example: A fraud model’s 99% accuracy may hide that it detects no fraudulent transactions when the data is highly imbalanced.
• Formal product review or step-by-step tutorial: None is presented.

Main Speaker and Sources

• Main speaker: An unnamed machine learning engineer and content creator.
• Sources and references mentioned: The speaker’s work experience; Reddit posts and replies from developers; an AI engineer’s post; and examples involving ChatGPT, Claude, Microsoft Copilot, Kaggle, scikit-learn, and PyTorch.