TECHNICAL AI SAFETY // FIELD MAP
← field map
UNIT 6 · START CONTRIBUTINGchapter 7 · reading

Next steps: Other fellowships

BlueDot Impact · Technical AI Safety · unit 6, chapter 7
Course pageNext steps: Other fellowships — BlueDot ↗
TL;DR — Chapters 6.5 and 6.6 assume you want to be an alignment researcher or a policy analyst; this one is for everyone who sorts into neither. It names two programs that select on transfer rather than on prior AI-safety credentials — Tarbell, which turns working journalists into AI reporters and places them in real newsrooms, and the Principles of Intelligence (formerly PIBBSS) Fellowship, which pays neuroscientists, physicists, ecologists and economists to spend three months aiming their own field's tools at AI safety — plus 80,000 Hours' longlist of 200+ fellowships for when neither fits. The thing to remember: an unusual background is an advantage at these programs and a liability at the mainstream ones, so route accordingly. The binding constraint is almost always a once-a-year deadline you missed by two weeks.

Unit 6 is a funnel, and by chapter 7 it has already sorted most readers — chapter 5 to MATS, ARENA and the lab residencies, chapter 6 to the policy fellowships. This chapter exists because that sort is lossy: the ecosystem needs reporters who can read a model card, epidemiologists who understand what a capability evaluation is, and control theorists who have never opened a transformer implementation but know a great deal about feedback under uncertainty. The course's answer to "what if I'm one of those people?" is short, and correctly so — the value here is three links, not an argument.

Why "other" is a real category and not a leftovers bin

Mainstream AI-safety fellowships select on proximity: how close are you already to doing the work — your GitHub, your ML coursework, your prior interp paper. That filter is efficient, and it systematically rejects the person whose contribution would be largest precisely because it is orthogonal. A macroeconomist who has spent a decade modelling principal-agent problems has something to say about gradual disempowerment that no amount of PyTorch fluency substitutes for.

The programs here invert the filter. They select on research ability or craft in some field plus demonstrated interest in AI risk, and explicitly do not require fluency in the safety literature on arrival. That is the structural fact worth internalising: applying to a transfer-selective program with a non-standard CV is a different game from applying to a proximity-selective one, and the same application does well at one and badly at the other.

The journalism lane: Tarbell

The Tarbell Fellowship is run by the Tarbell Center for AI Journalism and is the clearest example of the pattern.

"A one-year program for journalists interested in covering artificial intelligence."— The Tarbell Fellowship, Tarbell Center for AI Journalism (2026)

The structure is ten weeks of remote AI-journalism training, a week-long in-person summit in the Bay Area, then a nine-month placement inside a real newsroom — past hosts include Bloomberg, The Guardian, TIME, MIT Technology Review, The Verge, NPR, NBC News and Scientific American. It is paid at $60–80k for standard fellows and $90–110k for senior fellows with five or more years of experience. The Center also runs story grants ($1k–$20k for freelancers and staff reporters), senior residencies, and Transformer, its own editorially independent publication.

Two things a technical reader should note. First, the placement is the product — training is just the means of making you employable inside an existing newsroom, and the pitch is the alumni conversion rate to full-time journalism jobs (the course quotes 42%). Second, the cycle is annual and early: 2026 applications closed 7 January 2026, selections landed in April, and the programme runs June 2026 to May 2027. Reading this mid-year means you are looking at next January.

The transfer lane: Principles of Intelligence (ex-PIBBSS)

The Principles of Intelligence Fellowship — the organisation renamed from PIBBSS, and pibbss.ai now redirects to princint.ai — pairs researchers from adjacent disciplines with AI-safety mentors for roughly three months of full-time work on a project that bridges the two.

"A PhD is not required, but we are looking for something like PhD-level research ability or equivalent independent work experience."— PIBBSS Fellowship, Principles of Intelligence (2026)

Fields it explicitly welcomes: mathematics, neuroscience, philosophy, physics, political science, ecology. Cohorts are small (~20), the stipend is $3,000/month with accommodation, meals and a return flight covered, and the programme is built around retreats, a shared office, dedicated mentoring and a closing symposium. Alumni have landed at Anthropic, Google, Oxford, Harvard and the UK AI Safety Institute.

The course text is stale here — check the live page. BlueDot describes this as a "3-month summer programme with retreats in London or The Bay" with tracks in Cooperative AI and Gradual Disempowerment. As of the current cycle it runs November 2026 – February 2027, primarily in person in Cape Town, with four tracks: Gradual Disempowerment, ASI Safety via AIXI, Safe Pareto Improvement, and Corrigibility. The 2026–27 deadline was 20 July 2026. Treat every date and location in unit 6 as a pointer, not a fact — verify at the source before you plan around it.

When neither fits: search the longlist properly

The third resource is 80,000 Hours' Fellowships Longlist, an Airtable of 200+ programs. A database that size is only useful if you query it rather than read it, and the query that matters is not "AI safety" — it is your own field's name crossed with the calendar. Filter for programs that take your discipline, sort by deadline, and put every deadline in the next twelve months into a calendar with a four-week lead reminder. Almost every fellowship in this space is annual; missing one costs a year, and there is no waitlist.

The chapter also gestures at biosecurity as an adjacent problem without linking a program for it. That gap is worth filling yourself — the same "transfer" logic applies to pandemic preparedness, and BlueDot runs its own biosecurity course as an entry point.

Readings, linked

The course sets no time budget for this chapter — all three items are pages to skim, not papers to study. Read whichever of the first two matches your background; if neither does, go straight to the longlist.

Exercises

This chapter ships no exercises — it is a link list inside unit 6's "next steps" run. One is added below so the chapter still produces an artefact.

  1. Build your fellowship deadline board code — field map extra. Turn the longlist from something you read once into something that pages you. Pull every fellowship you are plausibly eligible for out of the 80,000 Hours Airtable plus aisafety.com/training, and produce a dated board of the next twelve months with a reminder four weeks before each deadline. What a good answer has: at least 10 programs, each with name, URL, eligibility in one line, deadline (or "rolling" / "cycle unknown — check in month X"), stipend, and a one-line honest note on your fit; at least three of the deadlines verified against the program's own site rather than the aggregator, since aggregators go stale exactly the way BlueDot's PrincInt description did. Start here: (1) export the Airtable view to CSV (share view → Download CSV) into fellowships.csv; (2) in a notebook, load it with pandas and filter rows whose field/eligibility columns match your discipline; (3) for each survivor, fetch the program page with httpx + selectolax and grep for date-shaped strings to spot deadlines the CSV lacks — expect this to half-work, and hand-check the rest; (4) normalise deadlines with dateutil.parser, dropping anything already past; (5) emit an .ics file with the ics package, one all-day event per deadline plus one four-weeks-prior alarm, and import it into your calendar; (6) re-run it quarterly. If you want the LLM shortcut, feed each fetched page to a cheap model with a strict JSON schema ({name, deadline, eligibility, stipend, location}) instead of writing extractors by hand — but keep step 3's manual verification, because a hallucinated deadline is worse than no deadline.

Go deeper

Next: Next steps: Programs · Back to the map.