Training
Training is the process of adjusting a model's parameters so its predictions get closer to the correct answers.
What Is Training?
Training is how a model goes from a random, useless starting state to something that produces useful predictions. It works by repeatedly showing the model examples from a dataset, checking how wrong its predictions are, and nudging its parameters slightly in the direction that reduces that error.
Input Example
From the DatasetOne example the model hasn't mastered yet.
Model Prediction
Current GuessStarts random and useless, improves over time.
Compare to Correct Answer
Measures ErrorHow wrong the prediction actually was.
Adjust Parameters
Small NudgeMoved in the direction that reduces error.
Repeat
Many Times OverThis cycle runs across the whole dataset.
Key Idea
Training happens once, ahead of time, over a large dataset. Inference happens every time the model is actually used, on new input it hasn't seen.
Why Training Is Expensive
- Repetition — a model typically sees the training data many times over, gradually refining its parameters each pass.
- Scale — modern models have billions of parameters, and every one of them gets adjusted on every training step.
- Compute — the parallel matrix computations training requires are why GPUs, not regular CPUs, are used for this stage.
- Data volume — foundation models are trained on enormous datasets, which itself takes significant time and infrastructure to process.
Common Mistakes
Confusing training with inference
Training happens ahead of time to produce a model; inference is using that already-trained model afterward — they have very different cost and time profiles.
Assuming every AI application involves training
Most applications built on top of an existing model — via prompting, RAG, or fine-tuning — never do full training themselves.
Assuming more training data always helps equally
Data quality and relevance matter as much as volume — a large but noisy or unrepresentative dataset can hurt more than help.
Interview Question
What is training, and why is it computationally expensive?
Training is the process of adjusting a model's parameters so its predictions get closer to correct answers, by repeatedly showing it examples, measuring how wrong it is, and nudging the parameters to reduce that error. It's expensive because modern models have billions of parameters that all get adjusted on every step, the model typically sees the training data many times over, and the underlying computation is the kind of parallel matrix math GPUs are built for. It happens once, ahead of time, to produce a model — which is different from inference, which is using that already-trained model on new input afterward.
What an interviewer may ask next
- Why does training require GPUs rather than regular CPUs?
- Does building an AI application always require training a model?
- Why doesn't more training data always improve a model?
Explain It in 30 Seconds
Training is the process of adjusting a model's parameters, over many repeated examples from a dataset, so its predictions get progressively closer to correct answers. It happens once, ahead of time, and is computationally expensive because of the sheer number of parameters and the volume of data involved — which is why it relies on GPUs. Most applications built on top of an existing model never do full training themselves; they use prompting, RAG, or fine-tuning instead.