Data Science in the Age of LLMs: How Foundation Models Are Replacing Traditional ML Pipelines
For a decade, the data science workflow was: collect data → clean data → engineer features → select model → tune hyperparameters → deploy. Foundation models are collapsing this entire pipeline.
1. The Old vs. New Data Science Workflow
| Step | Traditional ML | LLM-Powered |
|---|---|---|
| Data Labeling | Weeks of manual annotation | Few-shot examples in prompt |
| Feature Engineering | Domain expertise + pandas | Embeddings from pre-trained models |
| Model Selection | XGBoost vs. Random Forest vs. SVM | Choose foundation model + prompt |
| Training | Hours to days on GPUs | Zero — use pre-trained weights |
| Deployment | Flask/FastAPI + model serialization | API call to hosted model |
2. When Traditional ML Still Wins
Foundation models don't replace everything. Use traditional ML when:
# Traditional ML is better for:
# 1. Tabular data with clear feature relationships (use XGBoost)
# 2. Real-time scoring at < 1ms latency (use lightgbm)
# 3. Domains with limited text data (sensor readings, financial ticks)
import xgboost as xgb
model = xgb.XGBClassifier(n_estimators=500, max_depth=6)
model.fit(X_train, y_train) # Still faster and cheaper than LLM inference
The modern data scientist is a hybrid practitioner — knowing when to reach for a prompt versus when to reach for scikit-learn.
