Where the next generation of AI-assurance founders actually is, program by program: cohort size, cadence, whether the names are public, and exactly what to hook the sourcing engine to. Ends with a prioritised build order.
Sourceability is a judgement about whether we can get participant names (and ideally an identity anchor: LinkedIn, GitHub, personal site, arXiv author ID) programmatically and repeatably.
Rough field shape for calibration: BlueDot alone reports 7,000+ people trained since 2022 (bluedot.org), Apart reports 3,500+ sprint participants and 100+ research fellows (apartresearch.com/donate), SPAR reports 400+ mentees across cohorts. The funnel is wide at the top and extremely narrow at the founder end — which is why the incubators (Catalyze, Halcyon) are worth more per name than the courses.
| Program | What it is | Cohort + cadence | How to get names | Sourceability |
|---|---|---|---|---|
| MATS matsprogram.org | The flagship alignment research fellowship. 12-week in-person, Berkeley + London; scholars matched 1:1 with mentors from Anthropic, GDM, Redwood, UK AISI, MIRI. | 120 fellows / 100 mentors for Summer 2026, the largest to date; two cohorts a year (summer, winter) plus an extension phase. Acceptance ~5% [inference, from a community roadmap post, not an official figure]. | Public alumni directory at /alumni with one page per person (/alumni/<slug>) including current organisation and a bio. Symposium/demo-day write-ups posted to LessWrong. Individual streams announced per-mentor (e.g. Neel Nanda's winter stream). | Yes — scrape the alumni index and the per-person pages. Best single structured source in the field. |
| ARENA arena.education | 4–5 week ML/alignment engineering bootcamp at LISA, London. Curriculum descends from Redwood's MLAB. | Run eight times; ARENA 9.0 runs 5 Oct – 6 Nov 2026 at LISA. Roughly 2–3 iterations/yr. Cohort size not published [inference: order 30]. | No participant roster. Alumni surface downstream: ARENA states alumni become MATS scholars, LASR and Pivotal participants, and engineers at Apollo, METR, UK AISI. Apart co-hosts ARENA interpretability hackathons (ARENA 6.0, ARENA 4.0) whose project pages do carry team names. | Partial — mine via the Apart hackathon pages and via downstream MATS/Pivotal rosters. |
| Apart Research apartresearch.com | Sprint → Studio → Fellowship pipeline. Weekend research hackathons feed a 3–6 month remote Apart Lab Fellowship. Outputs land at NeurIPS, ACL, ICLR. | Hackathons roughly monthly (2026: AI Manipulation, Technical AI Governance, AI Control, AIxBio, Secure Program Synthesis, Global South, Secret Loyalties, Digital Minds, AI Incident Response). Individual sprints hit 500+ participants / 70+ projects. 3,500+ participants and 100+ fellows lifetime. | Full event index at /sprints/all; each sprint page lists submitted projects with author names. Verified: the site content-negotiates and will return clean text/markdown instead of HTML — trivially parseable. | Yes — highest volume, lowest scraping friction. Also the best early-signal source: hackathon winners are pre-credential. |
| SPAR (Kairos) sparai.org | Part-time, remote, 3-month mentored research. Technical + policy + security + biosecurity. Ends in a public Demo Day with a career fair (METR, Redwood, GovAI, MATS attend). | Two rounds/yr. Fall 2025: 80+ projects, plans to accept 200+ mentees. Spring 2026: 130+ projects (+~50%). Fall 2026: 240+ mentors, research 14 Sep – 14 Dec, Demo Day 19 Dec. 400+ mentees lifetime. | Mentor directory is fully public with per-mentor project pages (/projects/f26/?mentor=Name). Mentee names appear on the published-research list (co-authored arXiv/OpenReview papers) and at demoday.sparai.org. | Partial → yes. Mentors: scrape now. Mentees: scrape Demo Day + paper author lists after each December/June. |
| Astra Fellowship (Constellation) constellation.org/programs/astra | Fully funded 3–6 month in-person fellowship at Constellation's Berkeley centre. Empirical stream + strategy/governance stream. $8,400/mo stipend, ~$15K/mo compute per empirical fellow. | Annual; 2026–27 edition runs Sep 2026 – Feb 2027. Cohort size not published. >80% of the first cohort now work full-time in AI safety (Redwood, METR, Anthropic, OpenAI, GDM, CAISI, UK AISI). | No public roster. Constellation also runs short visiting-researcher stays. Names surface via co-authored papers, LinkedIn "Astra Fellow" self-labels, and the Redwood/METR alumni graph. | No → partial. Highest-quality cohort in the field; needs a relationship, not a scraper. Treat as a warm-intro target. |
| Pivotal Research Fellowship pivotal-research.org | London-based research fellowship, technical + policy tracks, AI safety and biosecurity. £8,000 stipend for the 2026 Q3 round. | Seven cohorts, 129 alumni as of the Apr 2026 call. Roughly 2 rounds/yr. | Public cohort pages with photo, project title and an explicit LinkedIn / GitHub / personal-site link per fellow — see /2024-fellows. Alumni have founded PRISM Evals, Catalyze Impact and Moirai. | Yes — and uniquely, the identity anchor is supplied for you. Founder-conversion rate is visibly high. |
| PIBBSS (Principles of Intelligence) princint.ai | ~3-month interdisciplinary fellowship pulling physicists, economists, philosophers, biologists into AI safety. 2026–27 edition in Cape Town; tracks on gradual disempowerment, AIXI, safe Pareto improvements, corrigibility. | ~20 fellows, Nov 2026 – Feb 2027, $3,000/mo + accommodation. 73 alumni to date (AISI UK, Anthropic, Oxford, Harvard, Google, FAR, Epoch, Simplex). | Public fellows & alumni page; cohort talks published on the PrincInt YouTube channel; closing symposium each spring. | Yes — small n, unusually senior and unusually weird. Good source of non-obvious technical founders. |
| AI Safety Camp aisafety.camp | 3-month online team-based program, Jan–Apr. Volunteer-led, very wide funnel, from mech-interp to advocacy. | Annual. AISC10 ran Jan–Apr 2025; AISC11 Jan–Apr 2026. Dozens of teams per edition, ~5–20 people per team. | The research-outputs pages list full team-member names for every project, plus outputs (arXiv, LessWrong, GitHub, MAISU lightning talks). One of the most complete public rosters anywhere. | Yes — scrape /research-outputs/aisc<N>-*. Noisier talent than MATS; excellent for breadth and for spotting repeat-participant clusters. |
| MLAB (Redwood Research) github.com/redwoodresearch/mlab | The original ML-for-Alignment Bootcamp (2021–22). 28 participants in the Jan 2022 iteration, co-run with Lightcone. | Historical. Superseded by ARENA (which explicitly draws on the MLAB curriculum) and by Redwood's work through Constellation/Astra. | No roster. Curriculum is open source. Value is as a historical alumni graph — MLAB alumni are now senior and are the mentors/angels layer, not the founder layer. | No — do not build a feed. Mine once, manually, for the senior network map. |
| Program | What it is | Cohort + cadence | How to get names | Sourceability |
|---|---|---|---|---|
| BlueDot Impact bluedot.org | The on-ramp. Technical AI Safety, AGI Strategy, biosecurity courses, plus project sprints. The single widest funnel in the field. | 7,000+ professionals trained since 2022, from frontier-lab staff to policymakers. Courses run near-continuously. | No cohort rosters. There is an alumni stories page (curated, a handful of names) and a community page. Slack/Discord community is members-only. | No — volume is huge but names are not public. This is a partnership play (sponsor a sprint, judge a demo day), not a scraping play. |
| SERI (Stanford Existential Risks Initiative) seri.stanford.edu | Stanford summer research fellowship + postdoc program in existential risk. Historically important as the original host of SERI MATS, before MATS spun out and became independent. | Annual summer fellowship; small. Much reduced relevance now that MATS is independent. | Stanford-side pages list fellowship structure; individual fellows surface through Stanford department pages and papers. | Partial — low priority. The MATS lineage matters more than SERI itself today. |
| GovAI Fellowships governance.ai/opportunities | Summer and Winter Fellowships (research track + applied track), Oxford/London, plus a DC winter fellowship. The main pipeline into AI governance. | Two cohorts/yr, ~3 months each. Alumni into UK/EU/US government, DeepMind, OpenAI, Anthropic, CSET, RAND, Oxford, Cambridge. | GovAI publishes blog posts describing each cohort's research topics, with fellow names attached; alumni also appear as GovAI report authors. | Partial — scrape the blog/publications index rather than hunting for a roster page. |
| Talos Network (EU policy) talosnetwork.org | Flagship European AI-policy fellowship: part-time European AI Policy Fundamentals programme, the Talos Summit in Brussels, then a full-time in-person Brussels placement. Also a senior Policy Leaders Programme. | 15–20 per cohort, twice a year — Spring cohort starts February, Autumn cohort starts August. Travel, accommodation and meals covered. | No public fellows roster (/fellows returns 404 as of Aug 2026 — verified). Fellows are placed at named Brussels institutions, so LinkedIn placement search is the practical route. | No → partial. Small n, high policy leverage, low technical-founder density. Manual. |
| Athena researchathena.org | 10-week remote mentorship program for women and marginalised genders in technical AI alignment, plus an in-person retreat. | Roughly annual cohorts. The Oxford in-person retreat had 25+ participants (fellows plus rotating speakers). | Announcement posts on LessWrong and the EA Forum; Manifund project pages carry retrospectives. No standing roster. | Partial — small but a deliberate correction to a demographically narrow founder pool. Worth a manual relationship. |
These convert researchers into founders. Per-name value is an order of magnitude above the courses.
| Program | What it is | Cohort + cadence | How to get names | Sourceability |
|---|---|---|---|---|
| Catalyze Impact catalyze-impact.org | Non-profit AI-safety org incubator, London (run out of LISA). Cofounder matching, seed funding circle, mentorship. For- and non-profit both supported. | Winter 2024/25 cohort produced 11 organisations; ~15 orgs incubated to date, majority having raised 6 or 7 figures. From 2026, multiple programs per year. | The cohort announcement (Introducing 11 New Ventures) publishes, per venture: co-founder names, website, contact email, location, legal structure, near-term plans and the exact funding ask. Wiser Human, TamperSec, Luthien, Lyra, More Light, Netholabs, Live Theory, Anchor Research, Aintelope, AI Leadership Collective, plus one stealth venture. | Yes — highest priority. This is a pre-seed pipeline published as a blog post. Watch /blog. |
| Halcyon Futures halcyonfutures.org | Grants, incubation and investment in AI safety, AI security and biosecurity founders. Explicitly recruits leaders out of business, policy and academia. | Rolling, not cohort-based. 25 projects launched, >$400M raised downstream; network of 1,000+ researchers, founders and policymakers. Also runs a MATS stream. | Request for Founders page states their thesis; portfolio/project names published on the site. Individual founders via portfolio pages and their MATS stream page. | Partial — but strategically this is a co-investor and referral partner as much as a source. Their RFF is a free read on where the gaps are. |
| Seldon Lab / def-acc / Entrepreneur First adjacency | Batch-based AI-safety startup studios. Appear repeatedly in alumni bios (e.g. a MATS alum founding WeaveMind at Seldon Lab Batch 2; TamperSec supported by EF def/acc and Impact Academy). | Batch cadence, small. | Batch announcements; alumni bios on MATS and Catalyze pages already name them. | Partial — pick these up as a side-effect of the MATS/Catalyze scrapes rather than as separate feeds. |
Programs give you cohorts twice a year. These give you a continuous signal, and they catch the people who never applied to anything.
| Watering hole | Why it matters | Access mechanism | Sourceability |
|---|---|---|---|
| LessWrong / Alignment Forum lesswrong.com · alignmentforum.org | Where research is announced before it is published, where program cohorts are announced, and where a new author's first high-karma post is the earliest credible talent signal in the field. | Public GraphQL API at lesswrong.com/graphql (explorer at /graphiql) returning posts with title, postedAt, baseScore, voteCount, commentsCount, user{username,slug}. Also RSS — /feed.xml?view=curated&karmaThreshold=30 (verified live). Same schema serves the EA Forum. | Yes — API. The best real-time feed available. |
| arXiv safety authors | cs.AI / cs.LG / cs.CY papers on interpretability, evals, control, red-teaming. Catches lab researchers and academics who never touch the fellowship circuit. | export.arxiv.org/api/query with search_query=cat:cs.AI AND submittedDate:[...]. Limits: 3 requests/second, max_results 30,000 in slices of ≤2,000, 503 on overrun (and 429s reported since early 2026 — back off). For affiliation and disambiguation, enrich via the Semantic Scholar API (author search + batch endpoints; 1 req/s with a key, or download the bulk datasets). | Yes — API, with rate-limit discipline. |
| NeurIPS / ICML / ICLR safety workshops | Accepted workshop papers are the highest-precision filter for "technically excellent AND working on safety" that exists. SPAR, Apart and AISC outputs all land here. | OpenReview API v2: client.get_all_notes(content={'venueid':'<Venue/ID>'}), then openreview.tools.get_profiles() on the author IDs. Profiles return full name, email domain, institutional affiliation, homepage URL and DBLP entry. | Yes — API, and it hands you the identity anchor and affiliation directly. Run three times a year at camera-ready. |
| 80,000 Hours job board jobs.80000hours.org | ~850 live roles. Not a source of names — a source of demand. Which orgs are hiring what tells you which sub-fields are capital-starved and which are overheating. | Per their own FAQ, all roles live in a publicly accessible Airtable view, ranked by an HN-style decay formula. Third parties already build filtered views off it. | Yes — but use it for market intelligence, not sourcing. |
| aisafety.com / aisafety.world maps aisafety.com/training · aisafety.world | The canonical registry of programs, orgs and events. Currently lists 15 upcoming training programs including several this page would otherwise miss: AIAF Fellowship, Iliad Intensive, ERA:AI, M3 Fellowship, AI Security Bootcamp, Gen Stream, Arcadia's AI Governance Taskforce, Horizon Fellowship, Lens Academy, FAS Policy Entrepreneurship. | The page exposes a public Airtable base view (including past programs) linked from the footer — airtable.com/appF8XfZUGXtfi40E/…. Use this as the crawler's seed list so new programs enrol themselves automatically. | Yes — this is the meta-feed. Build against it first so the engine never goes stale. |
| Twitter/X clusters | Interpretability, evals and AI-control researchers cluster tightly and announce jobs, papers and new orgs there before anywhere else. Follower-graph overlap with known MATS/Redwood/Apollo accounts is a decent proxy for field membership. | No free firehose since the API repricing. Practical options: maintain curated X Lists, monitor via a paid API tier, or derive the graph indirectly from arXiv/LessWrong identities. | Partial — expensive and brittle. Deprioritise; the same people are reachable through LessWrong and arXiv. |
Ordered by (names per unit effort) × (founder likelihood). Build 1–5 in the first sprint; they are all static-page scrapes or documented APIs.
| # | Feed | Hook | Refresh | Why first |
|---|---|---|---|---|
| 1 | aisafety.com training registry | Airtable public view (link in the page footer) → seed table of programs | Weekly | Meta-feed. Every other scraper registers against it, so new programs enrol themselves and the map never goes stale. |
| 2 | Catalyze Impact cohort posts | Scrape catalyze-impact.org/blog, parse venture blocks | Weekly (cohorts land 2–4×/yr) | Founder names + emails + funding asks, published. Straight into dealflow, zero enrichment needed. |
| 3 | MATS alumni directory | Scrape /alumni index → /alumni/<slug> detail pages | Monthly | ~120 fellows per cohort, each with current employer. The densest concentration of near-term technical founders in the world. |
| 4 | Apart sprints + projects | Scrape /sprints/all then each sprint page — request text/markdown, the site content-negotiates | Weekly | Highest volume and the earliest signal: hackathon winners are pre-credential, so you meet them before MATS does. |
| 5 | Pivotal cohort pages | Scrape /<year>-fellows and the current fellowship page | Per cohort (2×/yr) | Ships the LinkedIn/GitHub anchor per fellow. Demonstrated founder conversion (PRISM Evals, Catalyze, Moirai). |
| 6 | LessWrong / Alignment Forum | GraphQL API /graphql; RSS /feed.xml as fallback | Daily | Continuous signal. Alert on: first-time author clearing a karma threshold; any post titled "Announcing…". |
| 7 | AI Safety Camp outputs | Scrape /research-outputs/aisc<N>-* | Annual (May, after the edition closes) | Complete team rosters, dozens of projects, with outputs attached. Great for breadth and for spotting repeat collaborators. |
| 8 | PIBBSS fellows & alumni | Scrape princint.ai/about/fellows-alumni | 2×/yr | Small n (~20), senior, interdisciplinary. Non-obvious founders the ML-native funnel misses entirely. |
| 9 | SPAR mentors + Demo Day | Scrape /projects/f26/ (mentors) and demoday.sparai.org (mentees) | 2×/yr (June, December) | 240+ mentors is a ready-made expert/advisor network; Demo Day surfaces the mentee layer. |
| 10 | OpenReview safety workshops | API v2 — get_all_notes(venueid) + tools.get_profiles() | 3×/yr (NeurIPS, ICML, ICLR) | Highest-precision technical filter; returns affiliation and homepage without enrichment. |
| 11 | arXiv safety authors | API export.arxiv.org/api/query, ≤3 req/s + backoff; enrich via Semantic Scholar | Daily | Catches lab and academic researchers outside the fellowship circuit. Noisy — needs a keyword/venue classifier on top. |
| 12 | 80k job board | Airtable public view | Weekly | Demand-side intelligence: which safety sub-sectors are hiring, and therefore where the next companies are forming. |
| 13 | Constellation/Astra · GovAI · BlueDot · Talos · Athena · Halcyon | Manual — relationships, warm intros, sponsorship, judging slots | Ongoing | No public roster, so no scraper will help. These are the ones to invest relationship capital in — Astra especially, given >80% full-time placement. |
Names alone are worthless — the same person appears as a MATS alum, an Apart hackathon team member, an arXiv author and a SPAR mentor within eighteen months. Resolve on a stable anchor: LinkedIn URL (Pivotal gives it to you), GitHub handle, arXiv/Semantic Scholar author ID, OpenReview profile ID, LessWrong user slug. Store the graph, not the list. Repeat appearances across independent programs are the strongest single predictor worth ranking on [inference — plausible from the alumni bios but not something anyone has measured publicly].