TECHNICAL AI SAFETY // FIELD MAP
← field map
UNIT 6 · START CONTRIBUTINGchapter 8 · reading

Next steps: Programs

BlueDot Impact · Technical AI Safety · unit 6, chapter 8
Course pageNext steps: Programs — BlueDot ↗
TL;DR — The last page of the course is a list of four things BlueDot itself runs and pays for: a residential week to decide what you're doing (Context Week), micro-grants to unblock a project you've already started (Rapid Grants), salary-scale funding to go full-time (Career Transition Grants), and a five-day founder sprint with money attached (Incubator Week). Their published totals tell you what each one really is — Rapid Grants average about $2.3k across 679 awards, Career Transition Grants about $60k across 59. The thing to remember: the previous chapters listed doors other people control; this one lists the doors that open on evidence you can produce in a weekend.

Unit 6 spends most of its length on things you apply to: roles at labs, technical fellowships, policy fellowships, everything else. Each is gated by a committee comparing you to a pile of other applicants, on a cycle you don't set. This chapter closes the course with the other half of the picture — the funding side, where the gate is usually "is there a concrete thing you are doing, and is a few thousand dollars the reason it's stuck?" That is a far cheaper question to answer than "are you one of the eight people we're hiring," and it is the one most people finishing a course can actually answer.

Four doors, and what each is really testing

Context Week is the decision instrument. Four residential days in Berkeley, roughly twenty people, travel and accommodation covered, aimed at people with a serious professional or academic background who are weighing an AI safety career and can articulate what they still don't know. Prior safety experience is explicitly not required. What it buys you is the thing hardest to get from reading: enough contact with people already in the field to tell which of your career hypotheses survive.

Rapid Grants is the unblock instrument. Between $50 and $10,000, for compute and API credits, paid research access, event and meetup costs, travel, tooling — the small hard costs that stall a project mid-flight. It is deliberately low-friction: the application is minutes rather than hours, and BlueDot pays upfront by default rather than reimbursing.

"On average we reply within a day, and 9 in 10 applicants hear back within a week."— Rapid Grants, BlueDot Impact (2026)

Career Transition Grants is the runway instrument. You propose your own amount and duration for a defined stretch of full-time work on a transition into safety or biosecurity. There is no BlueDot-alumni requirement and no demand for an existing safety work sample — analogous work from another field counts. What it does demand is evidence of momentum and a plausible plan for what the funded period will produce.

Incubator Week is the founding instrument. Five all-expenses-paid days in San Francisco moving from threat model to intervention to pitch, with funding decided on the spot and up to $100k in grants behind a backed pitch. It is scoped for people technically serious about AI safety, biosecurity, cyber or catastrophic risk who are ready to leave what they're currently doing.

Read the totals, not the blurbs

The chapter prints two running counters, and they are the most informative thing on the page. Rapid Grants: $1,571,895 across 679 grants. Career Transition Grants: $3,513,485 across 59. Divide, and the marketing copy resolves into two completely different products. Rapid Grants runs at roughly $2,300 a grant — that is a compute bill, a conference flight, a year of an API key. It is not funding you; it is funding a line item, which is exactly why the turnaround can be a day. Career Transition Grants runs at roughly $59,500 — that is somebody's salary for a defined stretch, which is why the application takes about forty-five minutes and passes through evidence review and interviews.

The practical consequence is ordering. With an idea and no artifact, the move is not "apply for runway" — it is start the thing, hit the first hard cost, take the micro-grant, ship the artifact, then apply for runway with the artifact attached. Rapid Grants is explicit that it wants work already in motion. The two are stages of one pipeline, and people routinely apply to the second while still standing at the first.

Where this breaks

Three failure modes worth naming. First, the numbers on the page are running totals and the dates are cohort dates — both go stale the moment the course text is frozen, so treat every figure here as a snapshot and re-check the live program page before you plan around it. Second, secondary sources disagree with primary ones: BlueDot's own blog post announcing Rapid Grants describes a $50–$1,500 band with reimbursement after the fact and a BlueDot-participation requirement, while the live grants page says $50–$10,000, paid upfront, open to everyone with priority for community members. The live page wins; older write-ups of the same program are a trap. Third, and most important, a program is not a contribution. The counterfactual value in every one of these funnels is the work you do with the money and the week; the grant is leverage on an existing motion, not a substitute for having one.

The one thing to carry away: your cheapest next move is not an application to a job, it is a small artifact plus a $500 grant that makes it possible. In a field where hiring managers cannot calibrate on credentials, a published eval, a replicated result or a working probe is the only work sample that transfers — and Rapid Grants exists precisely to remove the excuse that you couldn't afford the compute.

Readings, linked

The course budgets no fixed reading time here — it's a link page. Open Rapid Grants first: it is the lowest-friction door and the only one that can be actioned this week.

Exercises

This chapter ships no exercises of its own — it is a link page closing the unit. Two are added here, clearly labelled as field map extras.

  1. The one-page transition memo (field map extra) — Write a single page, addressed to a grant committee, containing: the specific safety problem you would work on, the smallest artifact that would demonstrate you can work on it, what that artifact costs in money and weeks, and which of the four programs above is the right fit for that cost. Then map it against the actual eligibility text on the program page you picked. What a good answer has: a problem narrow enough to name a dataset or a model, an artifact you could finish in under a month, a cost figure with real line items (GPU-hours at a quoted rate, an API budget, a paywalled-paper bill), and an honest sentence on why the cheaper program is not sufficient — if you cannot write that sentence, you are applying one rung too high.
  2. Ship the artifact the grant would fund code (field map extra) — Build a small, publishable safety result end to end, on free compute, so that the grant application has something attached to it. Rapid Grants explicitly wants work already in motion; this is that motion. What a good answer has: a public repo or notebook, a stated hypothesis, a result that could have come out negative, and a paragraph on what you would do with 100× the compute — which is the paragraph that becomes the grant application. Start here: (1) pick one claim from an earlier unit that you can test at small scale — a refusal that survives paraphrase, a linear probe for a behavioural feature, a sandbagging check across prompt framings; (2) use a model that fits a free Colab T4 or a laptop: Qwen2.5-1.5B/3B-Instruct, Llama-3.2-1B/3B, or Gemma-2-2B via transformers, with nnsight or plain forward hooks if you need activations; (3) write the eval as fifty to a hundred hand-checked prompts in a JSONL file rather than reaching for a framework — inspect-ai is worth adding only once the harness outgrows a loop; (4) run it, plot the result with error bars over prompt variants, and note every place the effect vanishes; (5) publish the repo with a README stating the negative results as prominently as the positive ones; (6) apply for the compute you would need to run it on a frontier-scale open model, citing the repo.

Go deeper

That's the end of the course. Back to the map · or re-read the argument from the start at unit 1, chapter 1 — the failure modes in unit 1 read very differently once you know what unit 5 can and cannot measure.