When to Use Pre-trained Models

What you'll learn: You'll discover when it makes sense to use existing AI models rather than building your own from scratch.

The Core Idea

Imagine you need to hang a picture on your wall. You could spend months learning metallurgy, forging your own hammer, and crafting nails—or you could simply grab tools from the hardware store and finish in five minutes.

Pre-trained models are the "hardware store tools" of AI Engineering. They're models that others have already trained on massive datasets, often spending thousands of dollars and weeks of computation time. Instead of repeating all that work, you can download and use them directly.

When Pre-trained Models Make Sense

Use pre-trained models when:

• Your task is common — Need to classify images, translate text, or recognize speech? Someone's likely already built a great model for this.
• You lack resources — Training large models requires expensive GPUs and huge datasets. Most teams can't afford this from scratch.
• Speed matters — Remember the speed vs novelty trade-off? If you need results in days or weeks rather than months, pre-trained models align with engineering goals.
• Your data is limited — Models learn better when they start with knowledge. A pre-trained vision model already "understands" basic visual patterns—you just fine-tune it for your specific use case.

Train from scratch when:

• You're solving a truly unique problem
• You have specialized data that differs drastically from public datasets
• Your research goal demands understanding how models learn from the ground up

The AI Engineering Perspective

This directly ties back to your toolkit and engineering goals. You're optimizing for delivering working solutions, not publishing novel research papers. Pre-trained models let you skip 90% of the work and focus on the 10% that's unique to your problem.

Key Takeaway: Pre-trained models are your shortcut to production—use them unless you have a compelling reason to build from scratch.