Training
What is it?
Training is the process of teaching an AI model by showing it lots of data and adjusting its internal parameters until it performs well.
Explain like I'm 5
Why was it created?
Models start with random parameters and know nothing. Training is how they learn patterns from data to become useful.
Where is it used?
- Building new models
- Fine-tuning existing models
- Improving accuracy on a task
- Research and product development
Why should developers care?
Training determines what a model can do, and its cost and data choices shape AI capabilities, bias, and limits.
How does it work?
The model makes predictions on training data, measures its error with a loss function, and adjusts its parameters slightly to reduce that error. Repeating this over huge datasets gradually teaches it patterns.
Real-world example
To build a spam filter, you train a model on millions of emails labeled spam or not until it reliably predicts new ones.
Common use cases
- Creating models from scratch
- Fine-tuning for specific tasks
- Improving performance with more data
- Adapting to new domains
Advantages
- Produces capable models
- Learns directly from data
- Improves with quality data
- Enables task-specific models
Disadvantages
- Very expensive in compute and data
- Time-consuming
- Can bake in data biases
- Needs expertise and infrastructure
When should you use it?
When you need a model to learn a task from data, or to adapt an existing one (fine-tuning).
When should you avoid it?
When prompting or RAG with an existing model already meets your needs — far cheaper than training.
Alternatives
Related terms
Interview questions
Beginner
- What is training?
- How does it differ from inference?
Intermediate
- What is a loss function?
- Why does training need so much data and compute?
Senior
- When is training/fine-tuning worth it over prompting or RAG?
- How do biases enter during training?
Common misconceptions
- "You must train your own model to use AI" — most builders use pre-trained models with prompting or RAG instead.
- "Training happens continuously while using the model" — it's a separate phase from inference.
Fun facts
- Training large models can require enormous amounts of specialized hardware and energy.
- Most product teams never train a model from scratch — they reuse pre-trained ones.
Timeline
- 2010s — Large-scale training on GPUs drives the deep learning era
Learning resources
Quick summary
Training teaches a model by adjusting its parameters to reduce error over large datasets — costly, but how models gain their abilities.
Cheat sheet
- Teach a model from data
- Adjust weights to cut error
- Expensive: compute + data
- Often unnecessary vs prompting/RAG