Summary
"AI engineer" is one title hiding two very different jobs. One builds with LLMs — shipping applications, RAG systems, and agents on top of existing models. The other builds the LLMs themselves — training, fine-tuning, and researching model architecture. Both push AI forward, but they work in different ways, need different skills, and are very differently open to a new grad. For most people entering the field, the first type is where the jobs are and where you can start without a PhD.
Type 1: builds WITH LLMs
This is the application side. You take models that already exist and turn them into products and systems people use. It is the large majority of AI-engineering jobs.
- What they do: make AI apps that people actually want to use; connect a model to tools, APIs, and company data; build systems that keep working as usage grows; care about how the product feels and whether it helps anyone; get working features out the door.
- Core skills: prompt engineering, RAG and vector search, tools and integrations (MCP, APIs), system design, evaluation, and deployment/monitoring.
- The mindset: a product builder. Success is measured by whether the system is accurate, reliable, fast, and useful in production.
Type 2: builds LLMs
This is the model side. These engineers create the base models that the application side then builds on.
- What they do: invent new model architectures; run training and fine-tuning and make models cheaper and faster; spend their days on math, algorithms, and training data; squeeze more capability out of each model; move the state of the art.
- Core skills: ML theory, deep math, model architecture, pretraining and training pipelines, optimization, distributed training, data curation, and evaluation/benchmarking at scale.
- The mindset: a model creator/researcher. Success is measured by capability, efficiency, and pushing benchmarks forward.
Where they overlap
The line is not a wall. Skills like fine-tuning, quantization, and efficient inference sit in the middle — a Type 1 engineer who needs a smaller, faster, domain-tuned model touches Type 2 territory, and a Type 2 engineer ultimately wants their model to be usable in real applications. Both jobs count as AI engineering. They just start from opposite ends of the same stack.
Which to target as a new grad
For most people entering the field, Type 1 is the answer. There are far more application roles than research roles, and they are accessible without a PhD or deep ML-theory background — you can get hired on the strength of strong software engineering plus practical LLM skills (RAG, agents, evaluation, deployment) and a portfolio of working projects. Type 2 roles typically expect graduate-level ML, research experience, and comfort with heavy math and large-scale training infrastructure; they are fewer and more credential-gated.
That is not a knock on Type 2 — if you love the math and research, pursue it. But if your goal is to enter the field and have impact quickly, build with LLMs first. You can always move toward the model side later as you pick up ML depth.
Reading a job description
You can usually tell which type a role is from a few signals:
- Type 1 tells: "build LLM-powered features," RAG, agents, LangChain/LangGraph, API integration, "production," "ship," vector databases, deployment, evaluation of applications.
- Type 2 tells: "train/fine-tune models," PyTorch/CUDA, "model architecture," distributed training, RLHF, "research," publications, benchmarks, GPUs/clusters, "from scratch."
If a posting is heavy on product, integration, and shipping, it is Type 1. If it is heavy on training, math, and research, it is Type 2. Match your prep to the type you are aiming for instead of trying to be both at once.
Key takeaways
- "AI engineer" splits into two jobs: building with LLMs (applications) and building the LLMs themselves (model research/training).
- Type 1 (the majority of roles) needs prompt engineering, RAG, tools/MCP, system design, evaluation, and deployment; Type 2 needs ML theory, model architecture, training pipelines, and heavy math.
- They overlap at fine-tuning, quantization, and efficient inference — the line is a gradient, not a wall.
- New grads should target Type 1: far more jobs, accessible without a PhD, hireable on strong software skills plus practical LLM work and a project portfolio.
- Read job descriptions for tells — "RAG/agents/ship/production" signals Type 1; "train/fine-tune/architecture/research" signals Type 2 — and prep for the one you want.
What to do next
- Position yourself for Type 1 roles with The AI-era resume.
- Build proof with a Three-project portfolio.
- Follow the CS student AI-era roadmap or the Full-stack dev to AI engineer path.
Based on ideas shared in public industry posts (mostly LinkedIn, 2026). The ideas belong to their original authors; the wording and the extra detail here are ours.