TECHNICAL AI SAFETY // FIELD MAP
← field map
UNIT 6 · START CONTRIBUTINGchapter 5 · reading

Next steps: Technical fellowships

BlueDot Impact · Technical AI Safety · unit 6, chapter 5
Course pageNext steps: Technical fellowships — BlueDot ↗
TL;DR — The course closes by naming nine mentored research programs, from a 30-hour remote sprint to a 12-week funded residency in Berkeley. They differ in what you must already be able to do, what they hand you when you leave, and — the axis people ignore until it bites them — when their application window is open. Almost all select on a shipped artifact rather than a CV, so the highest-leverage move is to finish one small public project: that same artifact is the application material for all nine.

Every unit before this one taught you to see a problem — evaluations that miss what they claim to measure, interpretability that explains less than it appears to, control schemes resting on untested assumptions. This chapter is the course admitting that seeing a problem and being paid to work on it are separated by a gap reading cannot close. The field has few senior researchers and many capable people who want to work with them; the fellowship system is the market that clears between them.

What a fellowship buys, precisely

Three things, and different programs sell different mixes. Supervision — someone with better research taste than yours reading your work weekly and telling you which of your three ideas survives contact with data; that is the part you cannot self-serve, and the whole reason the format exists. An artifact — a paper, a workshop submission, a tool other people use. And a credential plus a network, which matters more in practice than on the merits, because hiring in a field this young runs on who has vouched for whom.

Notice what is not on that list: teaching. These programs assume you can already code. The bootcamp among them — ARENA — exists precisely because the others assume it, and positions itself as preparation for the research fellowships rather than a substitute.

Three axes that separate the nine

Commitment. SPAR and BlueDot's own sprint are part-time and remote — you keep your job. MATS, LASR, Astra, ERA and Pivotal require you to move to Berkeley, London or Cambridge for two to five months. That relocation, not the acceptance rate, is the real filter for most working engineers, which makes the part-time options a different product rather than a consolation prize.

Structure. LASR Labs puts you in a team of three or four aimed at one publishable paper. MATS, Astra, Pivotal and the Anthropic Fellows program pair you one-to-one with a mentor on their agenda. The team format reliably yields a paper with your name on it; the one-to-one format more often yields a job at the mentor's organisation, because the mentor has spent months evaluating you.

Selection. LASR wants evidence you have built something substantial. ARENA runs a coding test. MATS runs a two-stage process — apply to a broad track, then face work tests, coding assessments or writing samples inside a specific stream. The signal being read everywhere is can this person make concrete progress on an ambiguous problem, and the only compressed evidence for that is something you finished.

"All backgrounds and levels of experience are welcome and prior AI safety experience is not required."— MATS FAQ, MATS Research (2026)

Why "apply anyway" is right, and where it stops

It is right for a boring statistical reason: selection here is a matching process under noise. MATS literally runs a two-sided ranking match, where streams rank applicants and applicants rank streams, so a rejection is evidence about fit with one mentor in one cycle, not a verdict on you. Mironov's roadmap of these programs puts full-time on-site fellowships near 3–10% acceptance and part-time online ones near 20% — rates at which the second application's odds are barely changed by the first result.

Where it stops: an application with nothing behind it is a weak lottery ticket that also teaches you nothing. Since all nine read for the same signal, the dominant sequence is finish a small project → reuse it across every application → treat rejections as scheduling information. The other real constraint is visas: MATS provides J1s for non-US participants but notes J1 holders cannot participate part-time in the US, while the London programs advertise visa support.

"we urge you not to exclude yourself prematurely."— LASR Labs, LASR Labs (2026)

The part the course leaves out: the calendar

These programs are not continuously open and their windows barely overlap. As of early September 2026 several close within days of each other — MATS's Winter 2027 deadline is 6 September, ERA:AI's 13 September, LASR's 20 September, BlueDot's sprint 27 September — while Astra, Pivotal, ARENA, SPAR and the Anthropic Fellows program sit between cohorts, taking expressions of interest. Most on-site cohorts then run January to April, which is exactly why the deadlines pile up in September. Treat this chapter as a recurring calendar item, not a one-time decision, and check each program's own page: dates move, and the course page does not move with them.

The one thing to carry away: the application material for all nine programs is the same object — one finished, public, technical artifact with a writeup that says what you tried, what failed, and what you would do next. Build that first. Applying without it is where the "apply anyway" advice degrades into noise.

Readings, linked

The course sets no time budget here — these are nine program pages, not papers. Read them in the order below; the list runs roughly from highest bar to lowest, so if you are unsure where you sit, start at the bottom with SPAR or the BlueDot sprint and work up.

Exercises

This chapter ships no exercises of its own — it is a curated link list, and the course leaves the action implicit. Three field map extras below make it explicit.

  1. Field map extra — the two-page fit memo — Pick the three programs from the list above that you could actually say yes to, and write half a page on each: the format and dates, what its page says it selects for, the strongest piece of evidence you currently have for that criterion, and the gap. Then write one paragraph naming the single program you will apply to next and the date. What a good answer has: the gap stated as a missing artifact rather than a missing feeling ("I have never trained a model past a tutorial" beats "I don't feel ready"); at least one program you rejected for a concrete reason (cannot relocate, cannot take unpaid time, wrong track) rather than out of intimidation; and a date that comes from the program's own page, not from this one.
  2. Field map extra — the portfolio artifact code — Build the one object that serves as application material for all nine programs: a small, finished, public technical safety project with an honest writeup. Scope it to something you can complete in 20–30 hours; the sprint above is the supported version of exactly this. What a good answer has: a reproducible repo, a result you can state in one sentence, and a writeup whose most valuable section is what failed and why. Negative results are fine and are better evidence of research taste than a cherry-picked win. Start here: (1) pick one narrow claim from an earlier unit you can test — a published eval's score does not survive a prompt-format change; a probe trained on one distribution transfers poorly to another; a refusal behaviour is recoverable by a documented technique. (2) Choose a model that fits your hardware: gpt2-small or a 1–2B open model in a free Colab GPU session, or an API model if the question is behavioural rather than internal. (3) Use the standard stack — transformers plus transformer-lens or nnsight for internals, inspect-ai for evals, datasets for data. The ARENA curriculum is a free source of both scaffolding and project ideas. (4) Fix a baseline and a control condition before you look at results, and log every run. (5) Write the result up in under 1,500 words with the code linked. (6) Post it publicly and send it to one person who works in the area.
  3. Field map extra — the deadline tracker code — The failure mode this chapter invites is missing a window by a week. Build a tiny tracker so that never happens. What a good answer has: a single file listing each program with its URL, cohort dates, application open/close dates and your status, plus something that tells you about it before the date rather than after. Start here: (1) put the nine programs into a YAML or CSV file with a deadline field and a source_url. (2) Write a ~30-line Python script that sorts by deadline and prints anything inside the next 45 days. (3) Add a staleness check: flag any row whose dates you have not re-verified against its source page in 60 days, since these pages change and this one will not change with them. (4) Cross-check your rows against the community-maintained AISafety.com training directory, which tracks roughly 110 programs with open/closed status. (5) Run it from cron or a scheduled GitHub Action once a week.

Go deeper

Next: Next steps: Policy fellowships · Back to the map.