Constellation

Stated Theory of Change

Constellation's stated theory of change is the multiplier effect applied to AI safety talent. The causal chain:

1. Physical colocation of AI safety researchers, policy people, and operators in a single Berkeley hub creates serendipitous connections that wouldn't otherwise occur.
2. Structured programs (Astra Fellowship, Visiting Fellowship, Incubator) take promising people, expose them to experienced mentors, and accelerate their transition into full-time safety work.
3. Talent deployment places these people into frontier labs, policy organizations, and new safety orgs, multiplying the total amount of safety work being done.
4. Continuous operation (vs. episodic conferences) allows relationships to deepen over months, producing lasting collaborations and shared understanding.

The key claim: investing $1 in Constellation-style field-building produces more safety work than investing $1 in direct research, because each person who enters the field through Constellation produces years of downstream safety work. CG's Asya Bergal estimates the multiplier effect makes capacity-building "several times (maybe an order of magnitude) more impactful" than direct technical grantmaking.

Revealed Theory of Change

Constellation's actions paint a more specific picture than its stated mission:

What they actually optimize for is talent pipeline throughput. The Astra Fellowship is the core product: take ~20 promising people per cohort, connect them with senior mentors from frontier labs, and place them in safety roles. The 80%+ placement rate is the key metric. The Talent Mobilization Lead role ($170-220K salary) -- essentially a specialized recruiter for AI safety -- reveals this as the central function.

They are infrastructure for CG's capacity-building strategy, not an independent research institution. This is the most important divergence from the stated ToC. Constellation does not set AI safety research priorities, does not publish research, does not produce technical output. It produces people -- specifically, people who populate other organizations' research agendas. The theory of change is articulated by CG staff, the CEO search is announced by CG staff, and CG staff physically work from the space. Constellation functions more like a CG-funded operating program than an independent org with its own strategic vision.

The office space is both the mechanism and the product. Unlike research orgs whose output is papers, or policy orgs whose output is recommendations, Constellation's output is the space itself and the network effects it generates. This makes evaluation inherently difficult -- how do you measure the value of conversations that happened because two people sat near each other at lunch?

The incubator is aspirational, not demonstrated. No named organizations have graduated from it. The real incubation success (AI Futures Project) came from Astra, not the incubator specifically.

Key Assumptions

Assumption 1: Physical colocation in Berkeley produces safety-relevant collaboration that wouldn't otherwise happen.

• For: CG survey data shows 60% of GCR workers cite CG-funded programs as top-4 career influences. Named testimonials from Neel Nanda, Gabe Wu, and others credit Redwood/Constellation programs. The AI 2027 project explicitly emerged from Astra connections.
• Against: Oliver Habryka (who ran a comparable space) reported "few new projects from collaborations in the space, other than between people who already had a long history." Remote collaboration tools have improved dramatically. Berkeley is not the only place safety researchers work.
• Testable? Partially -- could compare publication rates and collaboration patterns of Constellation residents vs. similar researchers elsewhere.
• If wrong: Constellation becomes an expensive cafeteria and office rental service without meaningful safety impact.

Assumption 2: Growing the AI safety talent pipeline is net positive for safety outcomes.

• For: AI capabilities teams are growing rapidly (2-3x/year at frontier labs). Safety needs more people or the safety-to-capabilities ratio worsens. Every person who enters safety work through Constellation is one more person working on the problem.
• Against: Oliver Habryka's deep objection: the AI safety community may have counterfactually created the AGI arms race. If growing the field brings more people into proximity with capabilities work and some of them end up doing capabilities, the net effect could be negative. The Lightcone critique notes that "AI Alignment ideas/the EA community/the rationality community played a pretty substantial role in the founding of the three leading AGI labs."
• Testable? Extremely difficult. Counterfactual reasoning about whether individuals would have done capabilities work without safety exposure is speculative.
• If wrong: Constellation accelerates the very problem it's trying to solve by training people who end up at capabilities labs without sufficient safety impact.

Assumption 3: CG/OP can objectively evaluate Constellation's effectiveness despite deep personal and financial ties.

• For: CG has a professional grantmaking process with 11 staff on the capacity-building team. They conduct formal surveys. Their $150M+ annual allocation suggests serious analysis.
• Against: The 2023 critique documented multiple personal relationships between Redwood/Constellation leadership and OP grantmakers. CG staff work from Constellation's space. Nicole Ross moved from CG to Constellation. The CEO search was announced by CG. These ties make independent evaluation structurally difficult.
• Testable? Would require an external evaluation by a funder with no ties to Constellation.
• If wrong: $23M may be significantly more than a disinterested funder would allocate, and Constellation may persist in its current form regardless of effectiveness.

Assumption 4: Constellation can maintain organizational stability through continuous leadership turnover.

• For: The hiring of Meghna Mann (professional management background) and the active CEO search suggest an intentional professionalization. This could bring stronger operations than the founding team's EA-background approach.
• Against: All three co-founders departing within 2.5 years of independence is unusual. The org has never had stable long-term leadership. An active CEO search combined with a brand-new COO means both top positions are in flux simultaneously.
• If wrong: The org drifts without clear strategic direction, becoming a well-funded but purposeless hub.

Strengths

1. The talent pipeline is demonstrably effective. The 80%+ placement rate of Astra's first cohort is a real outcome. Named individuals at Anthropic, OpenAI, DeepMind, METR, AISI, and Redwood credit Constellation programs with shaping their careers.

2. The physical hub model fills a genuine gap. After Lightcone shut down, Constellation became the primary shared office for Berkeley AI safety. The 200+ weekly visitors suggest strong revealed-preference demand for the space.

3. Mentor quality is extraordinarily high. Jan Leike, Sam Bowman, Alec Radford, Trenton Bricken, Nicholas Carlini, Buck Shlegeris, Ryan Greenblatt, Lennart Heim -- this is a mentor list that few programs can match. Access to these people is a genuine, valuable, and scarce resource.

4. The AI 2027/AI Futures Project outcome is a genuine flagship success. A scenario read by 1M+ people, reviewed by Holden Karnofsky, Helen Toner, Miles Brundage, and Yoshua Bengio, and cited by policymakers -- this emerged from Constellation's ecosystem and represents real-world impact.

5. CG's strong institutional backing provides financial stability. $23M in committed funding and a funder who views capacity-building as "maybe an order of magnitude more impactful" than direct technical work means Constellation is unlikely to face funding crises in the near term.

Weaknesses and Risks

1. Near-total funder dependency. >90% of known funding from a single source (CG/OP). This is the most extreme funder concentration of any AI safety org I've reviewed. If CG rebalances its portfolio, shifts strategy, or faces its own funding constraints, Constellation has no fallback.

2. No independent public voice. Zero podcast appearances, interviews, or public talks by any leader. The theory of change is articulated by the funder, not by the org. This raises a fundamental question: does Constellation have its own strategic vision, or does it exist to execute CG's field-building strategy?

3. Governance opacity. Board composition unknown. No COI policy public. No annual report. Deep structural ties to the funder (personnel, physical colocation, personal relationships). For a $23M org in a small, trust-dependent ecosystem, this is a serious transparency deficit.

4. Continuous leadership instability. Three co-founders departing, active CEO search, brand-new COO -- the org has never had stable leadership for more than 12 months. Organizations this young typically need consistent leadership to establish culture and strategic direction.

5. The gatekeeper problem. Constellation decides who gets access to the most influential AI safety network in Berkeley. Selection criteria for hosting, fellowships, and incubation are opaque. The 2023 critique documented conformity pressure and social hierarchy concerns. This gatekeeping role creates power dynamics that warrant scrutiny and transparency.

6. Evaluation difficulty. Constellation's core output (network effects, serendipitous collaboration, researcher development) is inherently hard to measure. The 80% placement rate is a useful metric but doesn't distinguish Constellation's counterfactual contribution from general safety field growth.

7. The Habryka critique is not answered. If growing the AI safety community has counterfactually accelerated AI capabilities development, then Constellation's multiplier effect could multiply harm. This existential challenge to the field-building model has not been publicly engaged with by Constellation or CG in a rigorous way.

Cross-References

• MATS: Most direct comparison. Both run in Berkeley, share some mentors (Buck Shlegeris, Ryan Greenblatt), serve similar populations, claim similar placement rates (~80%). MATS is more research-focused (structured mentorship on specific research agendas), Constellation is broader (includes governance, security, strategy). MATS publishes more about its approach and outcomes. Ryan Kidd (MATS) has given public interviews; Constellation leaders have not.

• BlueDot Impact: Online courses serving a different segment (earlier in pipeline, broader audience, 4,500+ enrolled). Complementary rather than competing.

• 80,000 Hours: Career advisory. Constellation provides the physical infrastructure and direct fellowship experience that 80K Hours recommends people seek.

• Lightcone/Lighthaven: Former competitor that shut down, now runs events. Lightcone's shutdown created the gap Constellation fills. Habryka's philosophical critique of field-building is the strongest challenge to Constellation's premise.

• Redwood Research: Parent org from which Constellation separated. Redwood retained research (AI control); Constellation retained infrastructure and field-building. Now financially independent ($22K vs $12.6M revenue).

• CG/Open Philanthropy: Not a peer but effectively a controlling entity. CG's capacity-building strategy is the strategic framework within which Constellation operates. The relationship is closer to a subsidiary than to an independent grantee.

What Would Change This Assessment

Upward revisions if:

• Constellation publishes an impact report with counterfactual analysis (how many of the 80% placed would have entered safety work anyway?)
• A new CEO with a clear, publicly articulated strategic vision takes over and gives substantive interviews
• Board composition is disclosed showing genuine independence from CG/OP
• Named incubator graduates emerge as meaningful safety organizations
• An independent evaluator (not CG) assesses Constellation's effectiveness positively

Downward revisions if:

• CG reduces or conditions future funding, revealing that Constellation has no funder diversification strategy
• A second round of cultural criticisms emerges post-independence (the 2023 concerns were not addressed)
• The CEO search fails or produces another short-tenure leader
• Evidence emerges that Constellation's talent pipeline primarily feeds capabilities teams rather than safety-focused roles
• Comparable placement rates are achieved by much cheaper programs (e.g., online-only alternatives)

Self-Critique

What sources should I have checked but didn't?

• The full 990 PDF (not just the ProPublica extract) would reveal board composition, related entities, and grants given.
• LinkedIn profiles of current and former staff might reveal additional personnel flows and departure patterns.
• The Google Doc "Who should be Constellation's next CEO?" is dynamically loaded and couldn't be read, but likely contains crucial information about the leadership transition.

Where is this analysis potentially biased?

• I may be overly influenced by the 2023 Redwood critique, which is the most detailed critical source but is 3 years old and pre-dates Constellation's independence. Culture could have changed significantly under new leadership.
• The absence of podcasts/interviews creates an unfair framing -- I'm comparing Constellation's public communication to orgs like METR and MATS whose leaders are more media-present. Some effective organizations simply choose not to do public communication.
• I may be placing too much weight on the CG dependency concern. Plenty of nonprofits are funded primarily by one major donor and function effectively.

What would a thoughtful person who disagrees say? "Constellation doesn't need to be publicly visible to be effective. The value is in the network effects, which are inherently invisible. The 80% placement rate speaks for itself. The fact that CG is the primary funder isn't a problem -- it's a feature, because CG has deep context on AI safety and is the most competent allocator of field-building dollars. The leadership turnover is a healthy evolution from founder-mode to professional management. And the 2023 critique was about Redwood's culture, not Constellation's independent operation."

What's my single weakest claim? That Constellation lacks independent strategic vision. It's possible that Constellation has a very clear internal strategy that it simply doesn't communicate publicly. The absence of public voice might reflect a deliberate choice to invest in programs rather than PR, not a lack of direction.

What information would most change my view? A candid, long-form interview with James Bregan or the incoming CEO about: (1) what specifically Constellation does that no other org does, (2) how they measure success beyond placement rates, (3) what they would do differently if CG weren't their primary funder, and (4) how they've addressed the 2023 culture concerns. If such an interview existed, it would address the central uncertainty in this analysis.