Two years ago, building an AI product meant hiring a team that could curate datasets, engineer features, and train models from scratch. Today, most of that work has been replaced by three techniques that any competent software engineer can learn in a weekend.
That's the core argument in a recent breakdown by engineer Bryan Anthonio, drawing on Chip Huyen's AI Engineering (O'Reilly, 2025). The distinction he draws is clean: ML engineering is about building models. AI engineering is about adapting foundation models - large pre-trained models like GPT-4 or Claude that already handle a wide variety of tasks - to specific use cases.
Three Ways to Adapt Instead of Build
The article lays out the adaptation toolkit that has replaced traditional model training for most practical applications:
- Prompt engineering - writing better instructions and providing examples to guide model behavior. Zero infrastructure required.
- Retrieval-augmented generation (RAG) - feeding the model relevant documents from your own database so it can answer questions about your specific data, rather than relying only on its training.
- Fine-tuning - actually updating a model's internal parameters with your own curated dataset. This still requires some ML knowledge but is far less work than training from zero.
Each step up the ladder requires more resources but delivers more control over output quality.
The Hard Parts Haven't Gone Away
The barrier to starting an AI project has dropped dramatically, but two problems have gotten harder. First, evaluation: when your model can handle open-ended tasks like summarizing legal documents or writing marketing copy, defining "good output" is genuinely difficult. Traditional ML had clean metrics like accuracy and F1 scores. AI engineering often requires human judgment and custom rubrics.
Second, inference costs add up fast. Every query to a foundation model costs money, and optimizing that spend through techniques like distillation (training a smaller model to mimic a larger one) or quantization (reducing the precision of model calculations to save memory) is becoming its own specialty.
This shift matters for anyone building with AI tools, not just engineers. The companies making products you use daily - from Notion AI to ChatGPT plugins - are largely doing AI engineering, not ML engineering. They're assembling and adapting, not training from scratch. That's why new AI tools ship so fast now, and why the quality gap between a two-person startup and a Big Tech lab keeps narrowing.