Start here
Three ways in. Pick yours.
Engineers and researchers who want to do the work frontier labs hire for in applied AI and research: build systems around models, and make claims about them that hold up.
- book chapters
- 21
- courses
- 12
- recorded postings at 48 labs
- 1,827
Posting counts use the recorded board (Partial update · October 8, 2026); each entry keeps its own observation date and needs a fresh check. Explore opportunities.
Where are you now?
Choose your door
- Door 1
Finishing a degree, aiming at research
32recorded internships, fellowships and residencies; 50% of research-scientist postings mention publications
What you add here
Where it leads
Research internships · Fellowships and residencies
Begin: Chapter 1: what a model call computesWhy this door
Labs rarely hire new graduates into research. The way in is a program, mostly internships for enrolled students and a few fellowships open to any background, and a research record. What you add here: derivations and measured experiments you can extend into your own work, and the training skills research postings ask for.
It derives what a model computes and checks it against a real one, the habit a research record is built on.
- Door 2
A software engineer moving into AI
57%of research-engineer postings have no degree requirement stated; median base pay $413k
What you add here
Where it leads
Research engineer · Forward-deployed engineer
Begin: Chapter 1: what a model call computesWhy this door
Research engineer is the door without a PhD: most of its postings state no degree, and it pays what research scientists are paid. What you add here: how models work, how to measure them, and the training stack research teams ask for.
The book builds an agent from the first request up, and you run every step.
- Door 3
A PhD heading to a frontier lab
45%of research-scientist postings ask for PyTorch; 49% ask for post-training
What you add here
Where it leads
Research scientist · Post-training
Begin: Did model A really beat model B?Why this door
Your research counts, and some labs require it. What the postings add on top is engineering at scale: the language-model training stack, evaluations built as engineering, and production code. What you add here is that engineering, measured.
One run of an eval decides less than it looks like. This is the measurement a lab expects you to do first.
Every role has its own page, counted from the labs' postings: see all the roles. It is not for prompt tricks, a certificate, or a promise of a job. There is none of that here.
What you will be able to do
01
Derive it
The math behind the system, from first principles.
More
Work out the quantities behind the systems from first principles: what a model call computes, what prefill and decode cost, what pass@k means, how wide an interval is.
02
Build it
An agent from scratch, every part yours.
More
Build an agent from scratch: the loop, typed tools, memory and retrieval, durable state, and the isolation that keeps model output from becoming an action nobody granted.
03
Measure it
Error bars, paired comparisons, calibrated judges.
More
Measure models honestly: error bars, paired comparisons, calibrated judges, controls that can fail, and the right unit to count.
04
Find the failure
Locate the fault before you change anything.
More
Find where a system fails before changing it, and show that a fix changed what you meant it to change.
05
Read the numbers
Read a benchmark the way an interviewer will.
More
Read a benchmark, a lab's eval report or your own results with the questions an interviewer will ask: what was sampled, what was controlled, and what would prove it wrong.
The room
What is here
All of it is free to read; some courses ask you to register.
Measured studies: Calibrate an LLM Judge Before You Trust Its Score · Did Model A Really Beat Model B? Eval Variance, Measured · Measure Prompt Injection as an Attack-Success Rate, Then Bound It · Measure Agent Latency: Time to First Token, Decode and the p95 Tail
Pinned on the classroom wall
The promises every page keeps
1Every number has a source you can rerun
Measured results come with their code, data and a receipt, in a public repository.
2It runs on models you can run
Most experiments use small open models that fit a free Colab GPU, so you can repeat them without an API key. When a result needs a hosted model, the page names the model and what the run cost.
3The math is on the page
Derivations are written out, not waved at. Where a page measures something, the math says what to expect before the run does.
4What went wrong stays in
Failed controls, first drafts that were wrong, and results that cut against the argument are reported where you can see them.
5Written for people who read model output all day
Short, concrete and plain. If a sentence would make you suspect a model wrote it unchecked, it is a bug.
If a page breaks one, it is a bug. Tell Rod which page and what you found.
