BlueDot Impact

Theory of Change

BlueDot's stated theory of change is a three-step talent pipeline: find great people, equip them via cohort-based courses, and connect them to impactful roles. From their 2023 Impact Report: "We aspire for our brands to attract the most motivated, thoughtful and action-driven people in the world. We succeed in our mission by identifying people who could have a massive impact and supporting them to do so."

Their core bet is on a "great person" theory of AI safety: "Throughout history, a small number of people in the right positions at the right time have steered powerful technologies toward better outcomes. We find and prepare those people." They frame the problem as a workforce mismatch: ~2,000 people working full-time on AI safety versus a claimed need for 100,000 by 2030.

In mid-2025, they adopted a "defence-in-depth" framework organizing AI safety interventions into five layers (prevent, detect, constrain, withstand, adapt) and restructured their course portfolio accordingly. This broadened their scope well beyond the original alignment-focused training. In late 2025, they further expanded into a startup studio/incubator model, aiming to help founders create new organizations to address safety problems that existing orgs cannot.

What They Do

Courses: Comprehensive portfolio including AGI Strategy (25 hrs entry-level), Technical AI Safety (12 weeks), AI Governance (12 weeks), Biosecurity, Future of AI (2 hrs, open to all), Facilitator Training, and Technical Project Sprint. All part-time, online, cohort-based with expert facilitators. Free or pay-what-you-want.

Scale: 7,000+ people trained since 2022. Growth: 400 (2022) to 900 (2023) to 3,500 (2024). Target: 5,000-10,000 in 2025. ~100 student groups worldwide independently use their open-source curricula.

Impact data (2022 cohort only): 342 students surveyed. At course start, 18 (5%) were working on AI safety; two years later, 123 (37%) were. Of the 105 newcomers, 20 were full counterfactual career changes, 85 were accelerated transitions. Cost ~GBP 440/student. No comparable data exists for any later or non-alignment cohort.

Placements: Claims "hundreds" of alumni at Anthropic, DeepMind, UK AISI. Shares alumni data with partner organizations for hiring. No public breakdown by organization, role type, or seniority.

Operational quality: Published detailed retrospectives with quantitative data (completion rates, project quality scores, attendance tracking). 75% course completion rate vs. ~20% for typical MOOCs. June 2024 alignment retro: 50% project completion (up from 36%), 56% rated high-quality.

New ventures: $1M fund alongside AGI Strategy Course. Incubator weeks (v1 tested October 2025 with 7 participants -- honest admission that co-founder matching failed). Startup studio targeting pandemic PPE stockpiles. $130K in grants deployed so far. Rapid small grants ($50-$1,500) for project sprint participants.

Open source: All custom software on GitHub (16 repos). Curricula, session plans, facilitator training publicly available.

Key People

Dewi Erwan (Humphreys) -- Co-founder and CEO. Durham University (Engineering + IR). Previously Executive Director of EA Cambridge and Biosecurity Advisor to CERI. Community builder and operator, not a technical researcher. Thin public intellectual footprint -- no long-form interviews or detailed writing on AI safety strategy discoverable beyond a 2022 podcast summary.

Notable departures: Jamie Bernardi (co-founder) left February 2024 for GovAI, now in AI policy. Adam Jones (Technical AI Safety Lead) left for Anthropic mid-2024. Both departures were amicable and consistent with the career pipeline BlueDot claims to create. Jones was the most prolific intellectual contributor, and his departure leaves a gap in technical leadership.

Team: 7 people as of early 2026, scaling to 20. New hires include Sam Dower (Technical AI Safety, Oxford CS) and Luke Drago (Governance, co-authored "The Intelligence Curse").

Money and Incentives

Total funding: ~$34M raised since 2022 (self-reported). $32.06M from Coefficient Giving/Open Philanthropy across 5 grants. ~$2M from unknown other sources.

Funding trajectory:

Year	CG/OP Grant	Notes
2022	$2.0M	Seed grant
2023	$1.9M	Two GBP-denominated grants
2024	$2.5M	2-year, amount changed Feb 2025
2025	$25.6M	3-year general support

Funder concentration: ~94% from CG/OP. No SFF, EA Funds, Longview, or government grants identified. No funding diversification strategy visible. Original 2024-2026 fundraising target was $5M before receiving the $25.6M grant -- a 5x overshoot that fundamentally changed the organization's capacity and ambition.

Business model: Entirely grant-funded. Courses are free or pay-what-you-want. Zero earned revenue. No path to self-sustainability without continued philanthropic support.

Cost economics: Cost per student declining from ~GBP 1,200 (2023) to ~GBP 550 (2024). At 5,000 students/year and GBP 550/student, course delivery would cost ~$3.5M/year. The $8.5M/year budget implies substantial spending on incubator, grants, SF expansion, and team growth.

Compensation: Salaries benchmarked to "SF tech salaries" per careers page. For a 20-person team at SF rates, compensation alone would be $3-4M/year.

Financial accountability: Zero public financial reporting. Not a registered UK charity (fewer reporting obligations). US entity too new for 990 filings. First UK Companies House accounts due by March 2027. We cannot verify compensation levels, expense breakdowns, or overhead ratios.

Ecosystem position: Functions partly as a recruiting pipeline for AI labs and safety orgs. Shares alumni data with Anthropic, Constellation, GovAI for hiring. The funder (CG/OP) funds both BlueDot and the organizations that receive BlueDot alumni, creating a closed loop within the EA-funded ecosystem.

What Others Say

The "Why not just Bluedot" critique (March 2025) -- AI Safety Hungary organizer argues the "just send people to Bluedot" meme discourages local field-building. Provides acceptance rate data: Alignment 28%, Governance 47%, Economics of TAI 27%. Notes that BlueDot cannot possibly absorb all interested talent and local initiatives serve complementary functions. The author is explicitly not anti-BlueDot: "I think they are awesome and it's not easy to find an organisation that has done more for the field of AI safety than them."

The course-stacking problem (EA Forum) -- A post titled "We do not live by course alone" captures a recurring pattern: "after the tenth or twentieth person tells me 'I've taken BlueDot's AGI Strategy and the CAIS AI Safety course, I'm not sure whether I should do ARENA or apply to MATS...' I start to notice a pattern. THOSE ARE NOT YOUR ONLY CHOICES."

AI safety insularity (Transformer News, December 2025) -- Deep-dive on how the AI safety ecosystem including BlueDot is insular: funded by "a tight circle of EA-aligned donors," creating intellectual bubbles where "groupthink may be mistaken for truth." The piece does not single BlueDot out but identifies it as part of a broader pattern.

CG's own impact validation (EA Forum, 2026) -- Coefficient Giving's capacity-building team lead wrote that in CG's 2023 survey, "60% of respondents listed a capacity-building program or organization that our team was funding in their top four influences." AGISF (BlueDot's precursor) was specifically mentioned in several testimonials as influential in career trajectories of people now at DeepMind, Redwood Research, and OpenAI.

BlueDot's self-critique (June 2024 retro) -- "On motivation, we think we are doing an okay, but not excellent, job here. A common failure for us here is participants failing to grasp the impact of catastrophic risks."

What's Absent

No public financial reporting of any kind for an organization that has received $32M+. First UK accounts due 2027. No annual financial report. No way to verify how funds are being deployed.

No independent board members. Five officers, all mid-to-late 20s, all Cambridge-connected, three co-founders. No external governance for a $25M organization.

No updated impact data beyond the single 2022 alumni survey of 342 students. Despite training 7,000+, no quantitative outcome data for governance, biosecurity, or more recent alignment cohorts. The 37% conversion rate from the earliest, most self-selected cohort may not generalize.

No funding diversification. ~94% single-funder dependency with no visible plan for alternatives.

No detailed placement data. "Hundreds at Anthropic, DeepMind, UK AISI" is claimed without a breakdown by organization, role, or seniority, and without distinguishing people who took a free 2-hour course from people who completed the full 12-week program.

No CEO public intellectual presence. Dewi Erwan has no discoverable long-form writing, interviews, or detailed strategic thinking published since a 2022 podcast.