AI Researcher Path¶

In one line: How people become AI researchers, the difference between research scientist, research engineer and applied roles, and a realistic way to build a research record. Last reviewed: 2026-11-07

The short version¶

Research scientist roles set questions and write papers. Research engineer roles build the systems and experiments that make research run. Many labs value both. This split is a judgment from common practice, not a sourced taxonomy.
The usual signal is a track record: papers, reproduced results, open-source contributions or a thesis. Credentials help but a good public record can stand in.
A strong path for a self-learner: read papers well, reproduce one, then extend it with a small, honest experiment.
No salary, admission or hiring numbers are given because none were sourced.

Step 1: get the base¶

Maths for ML without fear, neural networks in plain English, transformer architecture.
The researcher path gives the learning order.

Step 2: learn to read papers¶

Start with the summaries in this collection, then read the originals. Attention Is All You Need and InstructGPT / RLHF are good first pairs.
For each paper, write: the claim, the evidence, what was not tested.
Use the ROADMAP and READING_GUIDE.

Step 3: reproduce something small¶

Pick a result you can run on one GPU: run Llama on a single GPU, fine-tune with LoRA.
Report what matched, what did not, and your seeds. Negative or partial reproductions are useful.

Step 4: contribute¶

Add an eval, a bug fix or a documented replication to an open-source project. Use set up an eval harness as a starting point.
Write a short technical note or blog post. Clarity is a research skill.

Step 5: choose a route¶

Route	Fit
PhD	Best when you want to set your own questions and are happy with a multi-year commitment.
Research engineer	Best when you like building systems; start from the ML engineer roadmap.
Applied research in industry	Best when you want measurable product impact; see the AI engineer roadmap.

Common mistakes¶

Reading only summaries. Read the paper, then check the code.
Claiming improvement from one seed or a tiny test set.
Skipping safety and evaluation work; both are active research areas. See safety.

Next steps¶

Build an AI portfolio for how to present research-style work.
AI careers map.

Sources¶

Judgment: the route comparison is editorial.
Paper details are in the linked summaries, each of which cites the original.