Session 0· 02· 10 min

ML vs LLM — what actually changed

What you'll learn
  • Contrast the classical ML workflow with the LLM workflow
  • Understand why LLMs skip feature engineering and task-specific training
  • Know when classical ML still wins

Classical machine learning and LLMs are both "AI" — but the developer workflow could not be more different. In classical ML you spend weeks collecting data, labelling it, tuning features, and training a model for ONE task. With an LLM you write a prompt, call an API, and read the answer.

Classical ML workflow

What you do for classical ML
Collect data
1,000s of rows
Label it
By hand, often
Feature eng.
Craft inputs
Train model
Hours–days
Evaluate
Accuracy, F1…
Deploy
Per task

LLM workflow

What you do with an LLM
Write prompt
Plain English
Call API
1 HTTP request
Read reply
Text in, text out
Iterate
Tweak the prompt
The shift
Classical ML builds a new model per task. An LLM is one giant pre-trained model that generalises to many tasks, controlled through natural-language prompts.
Classical ML
Task-specific
  • Needs labelled training data
  • Built for ONE task (spam? churn? fraud?)
  • Requires ML expertise
  • Deterministic, explainable (usually)
  • Deployable offline, cheap at inference
  • Months to build from scratch
LLM
General-purpose
  • Needs ZERO training data (uses pre-trained weights)
  • One model handles many tasks via prompts
  • Accessible to any developer
  • Non-deterministic, harder to explain
  • API call costs money per request
  • Hours to build a working prototype

Side-by-side capability matrix

 Classical MLLLM
Needs labelled training data
Works without domain experts
Handles free-form text input
Deterministic output
Good for tabular numbers
Good for unstructured text / vision
Explains its decisions
Runs offline, cheap at inference
Handles multi-step reasoning
When classical ML still wins
Need to score 50 million rows per day? Classical ML is orders of magnitude cheaper at that volume. Got a purely tabular dataset (sales forecast, credit risk)? A gradient boosted tree beats any LLM. LLMs shine on language, code, and unstructured data — not everywhere.
Key takeaway
LLMs are not a replacement for ML; they are a new tool that dramatically shortens the path from idea to working prototype for language-heavy tasks.