Executive summary: A digital marketing campaign run by Consultants for Impact achieved substantially stronger results than expected—generating over 11,000 newsletter subscribers, 44 million impressions, and 212+ career advising applications—suggesting that targeted paid social media can be effective for EA-adjacent orgs with defined audiences and clear offerings, though results may not generalize broadly.

Key points:

1. The campaign generated 11,000+ newsletter subscribers (5,500% year-over-year increase), 44 million impressions across Facebook, Instagram, and LinkedIn, and 212+ career advising applications, exceeding initial goals by approximately 900%.
2. Setting clear, specific SMART goals at the outset focused the campaign strategy; vague goals produce vague campaigns, and midstream goal changes are a leading cause of campaign failure.
3. The content strategy mixed three elements: memes for attention and shareability, valuable resources like CFI's free Giving Guide to build trust, and real stories of consultants who transitioned to high-impact work.
4. The campaign treated the effort as a test with a minimum three-month window (six months recommended with an agency), adopting a test-learn-repeat approach and adapting underperforming ads and posts rather than committing to a fixed plan upfront.
5. CFI's success depended on pre-existing conditions: a clearly defined target audience (management consultants), a strong website, established programming to convert interest, and a team willing to collaborate closely—conditions that marketing amplifies but cannot create from scratch.
6. For EA-adjacent orgs where reaching a specific population is the bottleneck to impact, paid social media is more accessible than commonly assumed, and the marginal cost of testing is low compared to the opportunity cost of never investigating it.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: While capability restraint—slowing AI development to ensure safety progress—faces significant practical challenges, especially internationally, it remains strategically important and potentially beneficial even in idealized scenarios, though advocates should acknowledge genuine trade-offs including concentrations of power, ceding competitive advantage, and prolonged background existential risks.

Key points:

1. The case for capability restraint rests on a basic logic: if safety progress takes time and unrestrained development risks human extinction or disempowerment in realistic scenarios, then significantly restraining AI development becomes necessary for survival.
2. AI development does not necessarily follow prisoner's dilemma incentives; depending on payoffs, it can resemble a stag hunt where mutual slow-downs are rationally preferred by all parties if they expect others to cooperate, creating multiple stable equilibria rather than forced defection.
3. Individual capability restraint (e.g., dropping out of the race or burning a lead) avoids requiring coordination but remains inadequate to address race dynamics, whereas collective restraint between multiple actors can be more effective but faces barriers around verifying compliance and restricting algorithmic progress.
4. Even in idealized scenarios with fully effective restraint and rational decision-making, the costs of delaying superintelligence's benefits can be significant; whether restraint is worthwhile depends on whether reductions in misalignment risk per unit of delay outweigh background risks of individual death and non-AI existential catastrophe during that period.
5. Compute-focused international governance appears promising because frontier AI relies on specialized, expensive, monitored infrastructure, but algorithmic progress is harder to restrict; at current rates, algorithmic improvements could allow a rogue actor with 10% of leading compute to reach parity within two years, potentially limiting effective pause duration.
6. Capability restraint could be net negative in multiple ways: by concentrating power in governance bodies or single actors, by ceding competitive advantage to authoritarian regimes, by prolonging background existential risks, and by exacerbating risks of great power conflict, implementation failure, and abuse.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: Given persistent expert disagreement about AI timelines, the author argues that adopting a broad distribution over when transformative AI will arrive—rather than committing to short or long timelines—is the epistemically humble and strategically sound approach, with implications for how individuals and communities should plan their work.

Key points:

1. The author defines transformative AI as a threshold where AI systems would be powerful enough to take over the world if misaligned or could double the rate of scientific and technological progress, and uses this to evaluate when timelines matter most for decision-relevant planning.
2. Expert forecasters disagree substantially on AI timelines, but the author notes that "long timelines have gotten crazy short" (shifting from 30+ years to 10-20 years) while "short timelines" now mean AI arriving within 2-5 years, with both camps updating on evidence.
3. Individual experts like Daniel Kokotaljo, despite being known as a short-timelines advocate, maintain broad distributions themselves (80% interval from 2027 to after 2050 for certain AI capabilities), and the broader expert community shows even greater overlap and uncertainty across forecasts.
4. The author recommends adopting a broad distribution over timelines rather than a single point estimate, noting that compressing uncertainty into one number obscures the fact that different time horizons (e.g., next presidential term vs. the one after that) represent "very different scenarios" requiring different hedging strategies.
5. In longer timelines (e.g., 2035 or beyond), the world will look substantially different due to geopolitical changes, technological shifts, possible AI-driven unemployment, and altered public sentiment about AI, which means approaches tailored to today's world may not work and new possibilities may emerge.
6. Long-term projects like founding organizations, building movements, writing books, and foundational research have high leverage in longer-timeline worlds and should not be ruled out; even though a book project has a 1-in-5 chance of arriving too late given the author's timelines, this leaves 80% of its expected value intact and addressing current neglect in AI safety creates additional value multipliers.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: SALA AI 2026 was an important Latin American AI event that brought together talented students, speakers, and safety-focused communities; the author describes valuable conversations with AI researchers and industry leaders about responsible AI development, and highlights a hackathon project on marine ecosystem analysis using machine learning.

Key points:

1. The author's community prepared for SALA by analyzing the International AI Safety Report 2026 to identify Latin American perspectives on AI risks and opportunities.
2. Apple is emphasizing responsible AI with focus on user data privacy, and limitations like poor generalization under distribution shift and weak calibration in high-stakes settings create real-world risks requiring worst-case robustness rather than average-case performance.
3. David Fleet identified deepfakes as a huge current challenge for the industry, with steganography being explored to identify artificially generated content, and emphasized that technology safety depends on both companies and responsible user behavior.
4. The concern that "situational awareness may allow AI models to produce different outputs depending on whether they are being evaluated or deployed" prompted Vincent Mai to share relatively simple evaluation techniques that can reveal behavioral patterns difficult to detect.
5. The hackathon team used pretrained models (Perch 2.0 and BirdNET) to extract embeddings from underwater acoustic recordings near the Galápagos Islands and applied clustering to identify structure in unlabeled marine soundscape data.
6. The team proposed developing a Kaggle-style competition to collaboratively build a labeled dataset for whale communication, received recognition from organizers, and aims to advance both the science and community engagement.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: While a recent study found that LLM access did not significantly improve novices' ability to complete dangerous biology tasks, measuring novice uplift is likely the wrong metric for assessing existential risk—expert uplift matters more and comes first, and future studies should focus on realistic threat actors and realistic threat scenarios.

Key points:

1. Active Site's randomized controlled trial found that 5.2% of the LLM group and 6.6% of the internet-only group completed a viral reverse genetics workflow, with no statistically significant difference (P = 0.759).
2. The author argues that novice uplift is probably the wrong frame for x-risk reasoning, because expert users will extract LLM capabilities before novices do, making novice uplift a late-stage lagging indicator rather than a leading one.
3. Historical threat actors like Aum Shinrikyo and the 2001 anthrax attackers were not novices; the more concerning threat model involves people with some domain expertise constrained by specific knowledge gaps, equipment access, or procedural bottlenecks—exactly the constraints LLMs are positioned to relieve.
4. Measuring expert uplift is methodologically challenging because experts are heterogeneous, but a within-subjects crossover design where each expert completes matched tasks with internet-only and LLM access, compared against themselves, could bypass this problem.
5. The study's experimental controls—blocking forum posting, communication tools, and restricting access to read-only internet—do not reflect realistic threat scenarios, and a better design would compare "internet plus all realistic tools plus LLMs" against "internet plus all realistic tools without LLMs" to isolate the model's marginal contribution while maintaining ecological validity.
6. The study tested frontier models with safety classifiers disabled, but a real threat actor would more likely download and fine-tune open-weight models, which represents a different threat surface; researchers should consider testing fine-tuned open-weight models through a bounded-capability adversary model that specifies constraints on compute, datasets, and domain expertise.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: The author argues that What We Owe The Future fails both as a justification for longtermism and as a persuasive work, mainly because its assumptions about influencing the far future, robustness of interventions, and key arguments about risk, values, and expected value are under-supported or implausible.

Key points:

1. The author argues we may not be able to predictably influence the far future due to limited information, cognitive limits, convergence dynamics, or chaotic effects like the butterfly effect.
2. The author claims there are no “robustly good” longtermist actions, since even interventions like clean energy could plausibly have large negative effects (e.g., enabling totalitarian lock-in or increasing wild animal suffering).
3. The author argues MacAskill’s framework relies on future humans being similar to us, which is unlikely given genetic, technological, and cultural changes.
4. The author claims that using expected value reasoning implies “strong longtermism” rather than longtermism, because far-future effects dominate near-term ones.
5. The author argues several specific claims in the book are under-supported, including high extinction risk, risks from stagnation, irreversibility of collapse, and the likelihood of value lock-in from AGI.
6. The author contends the book is unpersuasive to a general audience because arguments like comparing temporal to spatial distance and focusing on impacts “millions, billions, or even trillions of years” weaken intuitive appeal.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: The author argues that decision theory should not start from strong intuitions about what one should choose and then justify them, but instead should ground choices in independently compelling reasons, using verdict-level intuitions only to help discover those reasons.

Key points:

1. The author claims that a “verdict-level intuition” (a brute sense that one should choose a particular action) is not itself a reason, because such a verdict already presupposes that there are underlying reasons for that choice.
2. They argue that decision theory should proceed by identifying candidate reasons suggested by intuitions and then evaluating those reasons on their own merits, rather than treating intuitions as direct justification.
3. The author contends that reflective equilibrium, when interpreted as allowing mutual justification between intuitions and principles, still relies on the same mistaken use of verdict-level intuitions as justificatory.
4. In cases like Pascal’s mugging, the correct method is to assess reasons such as whether utility should be bounded, rather than inferring those reasons from the intuition not to pay.
5. The author argues that verdict-level intuitions are weak as predictors of unarticulated good reasons, especially in domains with poor feedback and where hard-to-articulate reasons are involved.
6. They suggest that this methodological point generalizes beyond decision theory to ethics and epistemology, where brute intuitions about conclusions should likewise be replaced with analysis of underlying reasons.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: The author argues that while biosecurity risks from AI, DNA synthesis, and weak institutions are real and in some cases growing, major human-targeting bioterrorism remains difficult and unlikely in the near term, with more plausible risks coming from institutional failures and agricultural attacks, and some optimism coming from detection systems and potential ML-enabled countermeasures.

Key points:

1. The author claims frontier LLMs currently provide limited practical uplift for novices in wet-lab virology (e.g., 5.2% vs. 6.6% task completion, P = 0.759), suggesting hands-on constraints remain a key bottleneck.
2. The author argues that biosecurity startups face a weak and volatile business case because government funding is inconsistent and may only scale after a catalyzing event, which historically tends to produce narrow, threat-specific spending.
3. The author claims DNA synthesis screening is fragile because it can be bypassed via short fragments, de novo or redesigned pathogens, and increasingly capable benchtop synthesizers, making the “chokepoint” assumption unreliable.
4. The author argues that creating and deploying human-targeting bioweapons is technically difficult, citing repeated failures by Aum Shinrikyo and limited effectiveness of non-state and some state programs, with success historically requiring massive state-scale infrastructure.
5. The author claims agricultural bioterrorism is much easier due to low biosafety requirements, simple deployment methods, weak detection incentives, and large economic impact (e.g., modeled $37B–$228B losses in U.S. scenarios).
6. The author argues current monitoring systems are mixed—wastewater surveillance shows promise for early detection, while systems like BioWatch have never successfully detected an attack—and that detection is limited by slow and uncoordinated response capacity.
7. The author speculates that machine learning may be more useful for rapid-response therapeutics (e.g., antibody design and mRNA delivery) than for offense, though this pipeline is currently incomplete and uncertain.
8. The author highlights pathogen-agnostic defenses like far-UVC and glycol vapors as potentially high-impact but underfunded public goods due to weak commercial incentives and limited evidence for large-scale deployment.
9. The author concludes that bioterrorism is a “low probability event” but worth preparing for, with the main bottlenecks being institutional and political rather than scientific.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: The author argues that animal activism is more effective when it targets environmental and institutional forces that shape behavior, rather than focusing primarily on individual persuasion.

Key points:

1. Differences in vegetarian rates (e.g., “40% of Indians” vs. “4% of Americans”) are attributed mainly to environmental factors like culture and availability rather than individual ethics.
2. Social systems (culture, government, institutions) shape behavior through “carrots and sticks,” norms, and by making some choices easier than others.
3. Institutional changes, such as NYC public schools’ Meatless Monday, can create large-scale dietary shifts (“the equivalent of creating 50,000 new vegetarians”) with relatively little direct persuasion.
4. Policies and norms are more durable than individual behavior change, which the author claims has an “84% recidivism rate” for vegetarianism.
5. When persuasion is necessary, the author argues activists should address environmental barriers and incentives rather than relying only on ethical or health arguments.
6. Providing structural support or alternatives (e.g., helping farmers transition or leveraging corporate pressure) is presented as more effective than moral exhortation alone.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Executive summary: Impact-focused programs require user buy-in and strong product-market fit in addition to sound theory of change; the author advocates treating user needs as a necessary (though not sufficient) condition for scaling cost-effective interventions, using lean product development practices to test and iterate.

Key points:

1. The author distinguishes between beneficiaries (those the program aims to help) and users (those who must be convinced to act on the program's behalf), arguing that strong user demand is often a necessary condition for theories of change to succeed in practice.
2. "Product-market-impact fit" combines user demand with cost-effectiveness by seeking programs that users genuinely want to engage with and that create cost-effective impact, a concept the author has refined through ~7 years in talent search and community building.
3. User needs and demand are faster to measure and optimize for than long-term impact metrics, making them useful leading indicators in early-stage development, though user demand alone is not sufficient for impact.
4. The lean product development process involves determining target customers, identifying underserved needs, defining a value proposition, creating a minimum viable product (MVP), and testing iteratively with real users to reduce uncertainty cheaply and quickly.
5. In the author's "Make your high-impact career pivot" bootcamp, extensive user research and piloting led to ~700 applications in the first month without paid marketing and 6 cohorts with 137 graduates in the first year, with likelihood-to-recommend scores of 7.9–9.1 out of 10.
6. The author acknowledges tensions between optimizing for user needs and maintaining cost-effectiveness, arguing these can be navigated by tying user needs clearly to impact and by treating user satisfaction as a precondition rather than an end goal.

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.