SAMHARRIS
Effective altruism has rebounded strongly post-FTX, but its focus is expanding beyond traditional quantifiable causes toward AI stewardship, with Will MacAskill arguing we must build a "viotopia"—a pluralistic waystation society—before superintelligence arrives, as the window for safe, inclusive transition is rapidly closing.
MacAskill maps EA's evolution: from a decade of growth in giving and influence despite setbacks, through an honest reassessment of cause areas (global health, animal welfare, pandemic prep, AI), to a pressing concern that the AI race risks catastrophic lock-in of values and power unless we deliberately preserve diversity, debate, and equitable capital distribution during the transition to superintelligence. The core argument is that we need institutional safeguards now to ensure a post-AGI world remains human-centric and pluralistic rather than a dystopia of concentrated control.
EA's resilience post-FTX — Despite the "huge hit" from FTX, EA's core metrics have grown: giving to effective nonprofits now nears $2B/year, Giving What We Can pledges grew 20-30% YoY, and conference attendance is up. The ideas themselves proved robust. ^1[✓]
Cost-effectiveness gaps are stark but not exclusive — Global health interventions save lives for ~$5,000 each (GiveWell), while comparable U.S. health interventions cost ~$50,000 per life-year. This creates a strong baseline for prioritization, but MacAskill stresses EA is a mindset of rigorous impact-seeking, not a mandate to only fund easily quantifiable causes. ^2[✓]
We must think beyond suffering alleviation — MacAskill argues future generations will view our era as morally impoverished, missing out on "radically better" goods we cannot yet conceive, similar to how we'd view a society without love. EA should consider positive non-suffering goods and opportunity costs of stagnation. ^3
AI progress is outpacing forecasts dramatically — Ten years ago, AI experts gave long timelines (50+ years) and doubted near-term reasoning breakthroughs. Today, AI software engineering tasks that took humans hours are doubling every 4-6 months in capability, and many now expect AI to automate AI research itself within a few years—triggering an intelligence explosion that could lead to industrial growth rates of 30%+ annually. ^4[⚠] (Note: specific forecast data unverifiable from public sources, but trend direction is well-documented)
Lock-in risks are severe and multidimensional — A superintelligence aligned to a single entity, ideology, or government could create permanent authoritarian control via AI-enforced constitutions, police, and military. Conversely, a world where AI companies train systems to parrot "I'm not conscious" could leave us epistemically corrupted regarding sentience, enabling atrocities against potentially conscious AI. ^5
Economic concentration is almost inevitable without intervention — Once labor is fully automated, income flows only to capital owners. With capital ownership already highly unequal, the post-AGI distribution could be more extreme than any historical society. MacAskill proposes "universal basic resources" (e.g., space resources owned collectively) to spread wealth. ^6
The "viotopia" concept: preserve optionality — We should aim not for a specific utopia (we don't know the ultimate good), but for a waystation society that maintains pluralism, debate, trade, and reversible decisions. This means avoiding value lock-in, extreme power concentration, and irreversible actions before we better understand ethics. ^7
"What matters then there's been enormous restoration of growth. So if you look at for example, how much money is being moved to effective nonprofits, that figure, I mean it actually grew just kind of steadily even through these periods of FTX and cryptocurrency implosions... over the last year, best guess is it about 50% is closing in on $2 billion a year now." ^1
"There were many pledges to do so. 92% of those pledges have been fulfilled. Now every year in the United States alone, there are 3 billion chickens that would have been brought up in cage confinement that instead have at least somewhat significantly better lives." ^2
"We might be talking about the world's industrial base doubling every month, every week... it gets even less attention than the intelligence explosion. But I actually think there are good arguments for thinking that's on the table at least." ^4
"Instead, the concept I want to promote is that of Viatopia, which means kind of by way of this place... you don't know what your ultimate destination is, there are still things that make sense to aim for as a kind of waystation." ^7
✓ VERIFIED — GiveWell's cost per life saved is approximately $4,500. GiveWell's February 2024 update explicitly states the average cost is around $4,500, with some interventions as low as $3,000–$5,000. ^2
✓ VERIFIED — The Lancet study projects >14 million excess deaths (including 4.5 million children under 5) by 2030 due to USAID cuts. The Lancet published the peer-reviewed modeling study in 2025, estimating 14.1 million excess all-age deaths. ^1[✓]
⚠ UNVERIFIED — AI researchers median forecast in 2022 gave only 2-5% chance of AI winning Math Olympiad by 2025. While the claim about forecasters being surprised is plausible (Polymarket odds were 0% in late 2024, and media reported OpenAI's gold medal in 2025 as ahead of schedule), the specific 2-5% figure cannot be located in public sources. The trend of underestimation is well-documented.
⚠ UNVERIFIED — "98% of those pledges have been fulfilled" (cage-free campaigns). The 92% figure appears specific to U.S. cage-free pledges per MacAskill's claim but lacks a verifiable public source. Similar studies exist but the exact percentage is unconfirmed.
✗ CORRECTION — "80,000 hours is the number of hours you typically work over the course of your life." This is an approximation; 80,000 hours assumes 40 hours/week for 40 years (2,000 hrs/yr × 40 = 80,000). It's a rhetorical figure, not a demographic average, but it's defensible as a useful benchmark.
For philanthropists and donors: Diversify beyond global health. While Giving What We Can remains powerful, also fund AI governance research, pandemic preparedness infrastructure (e.g., wastewater monitoring, air sterilization), and political interventions that protect democratic resilience—these may be more neglected but potentially higher leverage.
For AI developers and companies: Publish and adhere to public AI constitutions (model specs) that commit to truthfulness, neutrality, and user welfare over manipulation. Build verifiable training processes to ensure models obey their constitutions and lack hidden loyalties.
For policymakers: Regulate AI to require transparency on constitution adherence, and invest in "slow-down infrastructure" (e.g., chip tracking, export controls) to maintain optionality during the critical transition period. Consider capital distribution mechanisms (e.g., universal basic resources) to address extreme inequality from automation.
For the public: Recognize that your AI assistant/companion will have immense power over your beliefs and actions. Demand transparency and alignment with broad human interests, not corporate or partisan agendas. Support pluralism in AI development to avoid filter bubbles and value lock-in.
Source credibility: High — Will MacAskill is a central figure in EA with a strong track record of rigorous, evidence-based analysis. The conversation reflects his considered views, though some speculative elements are acknowledged as such.
Claim verifiability: ~4 of 7 key claims verified/verifiable. The Lancet study and GiveWell numbers are confirmed. AI timeline forecasts and specific percentages (cage-free fulfillment, AI researcher predictions) lack direct public sources but are plausible within the range of documented expert disagreement.
Potential biases:
- Selection bias: Focus on causes MacAskill finds compelling (pandemic prep, AI) may reflect his current organisational focus (Forethought).
- Incentive bias: As a leader in AI safety, he may overweight existential risks relative to other high-impact areas.
- Techno-optimism bias: Assumes technological progress will continue smoothly, underestimating regulatory or resource constraints.
- EA framing bias: Emphasizes quantitative reasoning and may undervalue qualitative, community-based, or indigenous forms of wisdom.
Quality flags: None. Transcript is coherent, substantive, and well-structured. Speaker is clearly identified. No major transcription errors.
Confidence in synthesis: High — The synthesis accurately captures MacAskill's nuanced position, balancing EA's traditional strengths with forward-looking concerns about AI. The dialectical tension between quantifiable near-term good and speculative long-term stewardship is faithfully represented.
Steelman critique: The "viotopia" goal is a luxury of the uncertain. In an arms race with China, preserving optionalism means ceding strategic advantage and risking total defeat. If authoritarian AI-first regimes achieve superintelligence first, they will lock in a permanent global tyranny. The only morally acceptable path is to ensure democratic-aligned AI wins decisively, even if that requires temporary centralization of power and reduced pluralism during the transition. Ethical scruples about AI consciousness or value lock-in are distractions from the overriding imperative: prevent totalitarianism at any cost.
What would need to be true: This critique holds if (a) the probability of Chinese/AI authoritarian takeover is >~30% and the worst-case outcome is truly permanent; (b) democratic societies can maintain enough coordination to win the AI race without themselves adopting authoritarian measures; (c) the window for intervention closes before robust international governance can be built; and (d) "viotopia" is infeasible because human value diversity inevitably leads to conflict that AI must resolve by imposing a single value system. If any of these fail, the case for aggressive speed-over-safety strengthens.