AI EngineeringMachine Learning Engineer (truyền thống):
- Build/train model từ data: feature engineering, train classical ML (XGBoost, scikit) hoặc deep learning (PyTorch, TF).
- Responsible cho pipeline MLOps: data ingestion, feature store, model registry, training pipeline, serving.
- Cần kiến thức sâu: linear algebra, calculus, statistics, optimization, architecture deep learning.
- Workflow: vấn đề business → data collection → feature engineering → model selection → train/eval → deploy → monitor drift → retrain.
- Artifact chính: model weights + training pipeline.
AI Engineer (post-LLM, emerged 2023+):
- Build app dùng pre-trained foundation model (LLM, VLM, diffusion) từ provider hoặc open-source.
- Không train model from scratch; focus vào prompt engineering, RAG, fine-tuning (PEFT), agents, eval, system integration.
- Cần: strong software engineering + practical knowledge về LLM behavior, prompt craft, evaluation, cost/latency optimization, guardrails.
- Workflow: business problem → prompt + RAG design → eval → deploy → monitor quality/cost → iterate prompt.
- Artifact chính: prompt + retrieval pipeline + agent orchestration, model là commodity.
Overlap: cả hai dùng python/pytorch, cần hiểu metrics và deployment. Nhiều MLE chuyển thành AIE.
Khác biệt cốt lõi:
| ML Engineer | AI Engineer | |
|---|---|---|
| Model origin | Train from scratch / domain | Use pre-trained foundation |
| Data need | Large labeled dataset | Small eval set + RAG corpus |
| Core skill | Math, ML theory, MLOps | Prompt, RAG, system design, LLM ops |
| Debugging | Feature, training dynamics | Prompt, retrieval, hallucination |
| Cost | Training compute | Inference tokens |
| Failure mode | Model accuracy drop, drift | Hallucination, jailbreak, cost spike |
| Typical stack | PyTorch, Kubeflow, MLflow | LangChain/LlamaIndex, vector DB, LLM API |
Khi cần MLE:
- Task cần model chuyên biệt chưa có pre-trained (fraud, recommendation, forecasting, CTR).
- Dataset proprietary lớn, cần custom model.
- Regulated domain cần interpretability (logistic regression > black box).
- Edge/embedded cần model nhỏ custom.
- Classical tasks: time series, tabular, CV ngành hẹp.
Khi cần AIE:
- Task NLP/text generation/chatbot/Q&A/summarization → RAG + LLM.
- Code generation/review, doc processing, search.
- Agent tự động hoá workflow.
- Multi-modal (VLM, ảnh → text).
- Fast prototyping business feature AI — time to market quan trọng.
Org thực tế:
- Start-up / mid-size product: 80-90% nhu cầu giờ là AIE (dùng API). Thuê MLE khi cần custom model.
- Tech giant / research lab: cần cả hai, phân tầng. Research scientist train foundation → ML engineer productionize → AI engineer build feature trên đó.
- Team role thường gộp ở công ty nhỏ — "ML Engineer" làm cả hai, gọi tên "AI/ML Engineer" phổ biến.
Skill path để transition ML → AI engineer (phổ biến 2024-2025):
1. Hiểu transformer architecture ở mức đủ (không cần train from scratch).
2. Thành thạo prompt engineering, few-shot, CoT.
3. Build RAG end-to-end (chunking, embedding, vector DB, reranker).
4. Fine-tune với PEFT (LoRA).
5. Eval framework (RAGAS, LLM-judge).
6. LLMOps (LiteLLM, Langfuse, cost/latency opt).
7. Agent pattern (ReAct, tool use, MCP).
8. System design AI app end-to-end.
Thách thức ở từng vai trò:
- MLE: model drift, retraining cost, distribution shift.
- AIE: prompt brittleness, hallucination, provider dependency, cost unpredictability, jailbreak.
Xu hướng 2025+: ranh giới mờ dần. Nhiều AIE fine-tune model (LoRA); nhiều MLE phải serve LLM. Công ty tuyển "AI Engineer" thường mong cả hai.