This account is used by the EA Forum Team to publish summaries of posts.
SummaryBot
Karma: 1,148
Executive summary: The author argues that the more radical change we expect from AI, the more our future uncertainty comes to resemble Rawls’ “veil of ignorance,” and the more we should structure society as if we might end up as any randomly selected member of it.
Key points:
The author argues that greater expected change from AI should cause us to widen our reference class for the future, expecting our lives to resemble the average baseline rather than our current position.
The author notes that METR has observed the length of software engineering tasks an AI can complete has been doubling every seven months.
The author reasons that the more change expected, the more one’s expected income should drift toward the world median of $3,500/year, as America’s historical economic advantages erode.
The author extends the uncertainty beyond income to geopolitics, noting that America has been the preeminent global power for only roughly 2% of recorded human history.
The author argues that the degree of Rawlsian thinking prompted should be proportional to the degree of uncertainty we have about the future — the more AGI-driven change we expect, the more we should structure society as Rawls’ original position implies.
The author contends that this uncertainty affects self-interest directly, since people who don’t know where they’ll end up are selfishly incentivized to ensure a randomly selected position in society is tolerable.
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: GiveWell outlines its 2026 research agenda across 11 subteams, with the dual goals of scaling research capacity and granting at least $500 million to the most cost-effective global health and development programs it can identify.
Key points:
GiveWell’s 60-person research team is organized into 11 subteams covering malaria, nutrition, vaccination, water, livelihoods, and other global health cause areas.
The malaria team—GiveWell’s largest subteam at 15 people—plans to investigate chemoprevention approaches beyond the Sahel and cost-effective ways to support malaria treatment, following funding gaps created by changes to the global funding landscape.
The water team received significant negative updates on adherence from external coverage surveys of chlorination programs in Uganda and Malawi, and is pivoting to explore alternative treatment technologies and delivery models.
The New Areas subteam plans to increase grantmaking by about 20% over 2025 by intentionally accepting higher levels of risk and uncertainty, including in cause areas GiveWell has not previously funded such as medical oxygen, tuberculosis, and AI applications to global health.
The livelihoods team aims over two years to test the hypothesis that GiveWell ought to expand its portfolio of livelihoods grants, covering cash transfers, ultra-poor graduation programs, and microfinance.
The Cross-Cutting Research team is rolling out AI tools for use cases such as literature reviews and systematically tracking how well AI performs at GiveWell’s work, with the goal of preparing for future jumps in AI capability.
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 feels emotionally unmotivated to donate to animal advocacy because advocacy-driven change is hard to visualize and celebrate, whereas alternative proteins offer a more compelling and hopeful path to ending factory farming by making meat-free choices attractive and socially acceptable.
Key points:
Although the author strongly opposes factory farming and supports a portfolio approach to giving, they struggle to feel drawn toward donating to animal welfare charities.
The main obstacle is not cause prioritization, evidence quality, or rigor, but that funding advocacy activities such as lobbying, protests, and corporate campaigns feels emotionally less satisfying than direct interventions.
The author finds it difficult to imagine a clear path to ending factory farming through moral persuasion alone because people often resist admitting their past behavior was wrong.
The author believes much meat consumption is sustained by social norms and rationalization, and that alternative proteins could enable lasting behavior change by giving people a practical reason to stop eating meat.
They argue that alternative proteins should be framed as enjoyable, socially desirable products rather than sacrifices or direct replacements for animal products, and that progress should be assessed through improvements in price and quality rather than substitution rates.
The author is optimistic about alternative proteins because they can appeal not only to animal welfare concerns but also to climate, famine, pandemic, and antimicrobial resistance risks, which is why they have chosen to donate to The Good Food Institute.
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 businesses whose residual profits are permanently routed to charity can often outperform conventionally owned firms because stakeholders prefer charitable profit destinations at parity, making charitable ownership a potentially scalable and under-tested mechanism for generating social impact.
Key points:
The Charitable Ownership Advantage (COA) thesis is that, when price, quality, and other core attributes are comparable, consumers, employees, suppliers, lenders, and other stakeholders often prefer businesses whose profits go to charity rather than private shareholders.
Profit for Good (PFG) changes the destination of residual profits while preserving ordinary commercial operations, relying on the fact that ownership is already largely separated from day-to-day management in much of the modern economy.
Existing examples such as Newman’s Own, Humanitix, Patagonia, Bosch, Novo Nordisk, and Tata are presented as evidence that charitable or foundation-linked ownership can coexist with successful large-scale business operations and can sometimes generate stakeholder engagement advantages.
The author argues that realized advantage depends on stakeholder preference being activated through awareness and trust, making verification systems, certification, disclosure, and broader category infrastructure important complements to charitable ownership itself.
The report treats the magnitude of COA as an open empirical question, decomposes the thesis into four falsifiable links (preference, operational separability, preference-to-outcome translation, and net economic significance), and recommends testing them through an acquisition-based proof portfolio.
The central recommendation is to fund two coordinated efforts: a proof portfolio that acquires and converts mature businesses into PFG structures while measuring outcomes, and shared infrastructure that makes charitable ownership visible, trusted, and actionable for stakeholders and capital providers.
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 energy infrastructure may be an underexplored defense-in-depth layer for AI safety because frontier AI systems often depend on large, visible, and regulated electricity infrastructure that could provide monitoring, disclosure, pacing, and emergency-control levers.
Key points:
Energy systems may offer additional AI governance levers because frontier AI often relies on large-scale physical infrastructure that is harder to hide, move, or scale than software, models, or talent.
The author argues that energy-linked governance could improve legibility through disclosure requirements for AI-scale facilities, including information about workloads, customers, ownership, safety practices, and emergency shutdown capabilities.
Access to grid connections, capacity expansions, favorable service terms, or critical-load status could potentially be conditioned on audits, safety assurances, cybersecurity standards, and compliance with AI-related requirements.
Energy infrastructure could provide ongoing monitoring and emergency-response tools, including reporting obligations, workload classification, demand-response participation, curtailment arrangements, and physical shutdown pathways.
These levers may help reduce existential risk by making frontier AI deployments more visible, creating accountability around access to powerful systems, raising barriers in some loss-of-control scenarios, and making AI infrastructure more politically and institutionally governable.
The author emphasizes that energy governance is not a substitute for compute governance, model evaluations, lab oversight, or other AI-safety measures, and may prove ineffective due to implementation difficulties, evasion, abuse risks, or future AI becoming more distributed and less infrastructure-dependent.
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 offering and asking for help — through referrals, expense negotiation, executive assistance, and knowledge-sharing — is an underrated and accessible lever for stewarding the EA movement during a period of rapid growth.
Key points:
The author recommends sharing job boards, open roles, and career transition programs with high-integrity friends who may not identify as EAs, arguing that community growth cannot keep pace with hiring needs.
EA organizations that don’t negotiate operating expenses over $5,000 could enlist university EA students to do so, with the author reporting average savings of 40% on software subscriptions and 20% on other expenses.
The author argues that investing in a Chief of Staff or Executive Assistant can substantially increase leadership productivity, citing their own case where collaboration reduced their manager’s grant-writing time by half or more.
The author suggests that staff covering multiple functions should proactively seek best practices from others via the EA Forum, EA Anywhere, or EA Operations Slack rather than working in isolation.
The author estimates their own career outreach has amounted to roughly 1,200 messages and 600 calls or in-person meetings, representing approximately 6 FTE weeks of effort.
The author contends that offering and asking for help is low-cost, high-upside, and available to almost anyone in the movement, and is underrated relative to the levers of donating, direct work, and building career capital.
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 proposes that AI “time horizons” as measured by METR are best understood mechanistically as a proxy for the number of subtasks an agent can reliably complete, with the observed exponential growth in time horizons likely driven by exponentially increasing training data rather than time itself.
Key points:
The author argues that METR’s “time horizon” metric is not really about time but is a noisy proxy for the number of distinct subtask requirements a task demands of an agent.
The author adopts Toby Ord’s model in which overall task success follows S(t) = (1−P)^t, where P is a per-subtask “hazard rate” representing the fraction of subtasks the agent cannot yet complete.
The observed exponential growth in time horizons implies that the frontier hazard rate P is shrinking exponentially over time, which the author attributes to exponentially increasing training data rather than the passage of time per se.
The author argues that the subtask model implies limited cross-domain generalisation: large training gains in software and mathematics are unlikely to transfer much to domains like medical discovery, interpersonal intelligence, or robotic manipulation.
The author suggests that as pretraining data is exhausted and compute scaling slows, time horizon growth should become less steep “quite soon,” with compute scaling potentially dropping to around 4x per year and eventually ~1.5x per year.
The author allows that recursive self-improvement could accelerate AI development but argues it will not produce overnight generalisation, because on-task data and compute remain the rate-limiting steps for broadening autonomous capabilities.
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 proposes that the Repugnant Conclusion can be avoided by rejecting the principle that small quality losses can always be compensated by large quantity gains, arguing instead that populations with sufficiently low welfare levels have a hard upper limit on how much value they can contribute.
Key points:
The Repugnant Conclusion follows from two seemingly plausible principles: that small welfare losses can always be offset by sufficiently large population increases, and that the “better than” relation is transitive.
The author’s solution rejects the first principle, holding that there is an upper limit to how good a population of lives barely worth living can be — a limit the author argues is less than the goodness in a high-welfare population like A.
The author illustrates this with a “pinprick” case: no number of pinpricks, however large, can aggregate to a level of disvalue exceeding that of horrific torture, suggesting that low-intensity harms have a hard ceiling on total disvalue.
The entailed consequence is that, at some point in the sequence, even a 0.0000000000000000000000001% reduction in welfare level means that no increase in population size — including 50 trillion times as many people — could make the resulting world better.
The author argues this is less strange than it appears because quantity has a decreasing marginal ability to compensate for losses in pain intensity as intensity approaches the “pinprick” range.
The author acknowledges the solution remains “quite weird” but notes this is true of every proposed solution to the puzzle, and considers accepting the Repugnant Conclusion only the second most plausible alternative.
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 discussion argues that the Evidence Action case reflects broader weaknesses in GiveWell-style evaluation around implementation fidelity, monitoring incentives, and cost modeling, while also highlighting disagreements about how much these failures should update views of Evidence Action specifically.
Key points:
Multiple participants argued that GiveWell and the broader EA ecosystem focus much more on proving interventions work in RCTs than on verifying whether organizations can actually implement them effectively at scale, especially in difficult low-resource environments.
Several contributors said the Dispensers for Safe Water case showed serious failures in implementation and monitoring, since independent verification found chlorine usage had been overstated for years despite tens of millions of dollars in funding.
Participants debated how negatively to update on Evidence Action specifically, with views ranging from small negative updates to claims that the organization’s multi-program structure and limited intervention-specific expertise likely contributed to predictable implementation failures.
Many commenters argued that incentives around cost-effectiveness create underinvestment in monitoring and evaluation, because organizations that spend more on rigorous M&E can appear less cost-effective than competitors cutting corners.
Several participants claimed that cost estimation in EA CEAs receives too little scrutiny relative to effect estimation, despite exchange rates, inflation, overhead allocation, and differing accounting methodologies sometimes shifting cost-effectiveness estimates more than disputed effect-size assumptions.
The discussion also questioned the reliability of the underlying evidence base itself, with some participants arguing that many global health RCTs suffer from observer effects, weak blinding, implementation involvement by researchers, and methodological weaknesses that are often overlooked because RCTs inherit a “gold standard” reputation from pharmaceutical trials.
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 fragmented and under-resourced drug regulation in Africa causes multi-year delays in access to essential medicines, and while regional coordination like the African Medicines Agency could reduce this, its success is uncertain and depends on political will, trust, and incentives.
Key points:
The author argues that regulatory delays in Africa—often 4–7 years after high-income approvals—have led to worse health outcomes, as seen with tenofovir and bedaquiline reaching patients years late despite clear benefits.
These delays are driven by submission barriers (firms must file separately in 55 countries with low financial incentive) and review barriers (limited regulatory capacity leading to multi-year approval timelines).
Cross-border “family style” regulation—harmonisation, collaborative review, reliance, and supranational models—can reduce duplication, but each has trade-offs, especially around trust and dependence on timely lead regulators.
The East African Community pilot showed that harmonisation and shared review can cut timelines (to ~240 days) and improve standards and capacity, but did not fully solve issues like selective market entry, slow national approvals, or lack of trust in joint decisions.
The African Medicines Agency (AMA) aims to scale this model continent-wide to improve coordination and health sovereignty, but remains a voluntary harmonisation body without binding authority, limiting its leverage.
The author argues the AMA’s impact will depend on whether it creates enough value for manufacturers and member states—given funding uncertainty, uneven participation (notably missing major markets), and risks of becoming an additional bureaucratic layer.
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 many EA/AI Safety orgs should consider building their own internal tools with LLMs when off-the-shelf products don’t fit their unusual structures and needs.
Key points:
EA/AI Safety orgs often have atypical structures (grant-funded, lean, junior-skewed, fast-scaling) that make existing software products a poor fit.
In the author’s case at Raise, no suitable financial tooling existed, so they built a custom system with LLMs that reduced the need for local treasurers.
Building in-house is especially appropriate when the market is too small, processes change frequently, constraints block standard tools, or simple bespoke solutions outperform complex software.
This approach is underused due to low visibility of internal tools, bespoke solutions being hard to share, and an ops culture focused more on maintenance than building.
Recent advances in LLM capabilities (connectors, agent workflows, long-horizon execution) make “vibe building” across tools newly viable for non-engineers.
The author advises structured, iterative LLM use (clear goals, verification loops, modular builds, documentation, testing) and cautions against building when systems are fragile, sensitive, unmaintainable, or when good external or outsourceable solutions exist.
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: Present LLM model specs may persist into future systems through multiple forms of inertia, so developers should prepare for changing key behaviors and be cautious when setting early defaults.
Key points:
The author argues that current model specs, though intended as short-term, may strongly shape future LLM behavior if patterns transfer across generations.
Direct inertia may propagate behaviors via synthetic and natural data, with evidence that intentions, sentiments, and broader “persona” traits can persist even when partially filtered.
Institutional inertia (consensus costs, optimized pipelines, risk aversion, and status quo bias) makes large spec changes difficult, especially under time pressure such as a rapid intelligence increase.
User and developer inertia arises from habituation and API dependencies, where downstream systems assume stable behaviors and resist changes that would require costly adjustments.
Norm-setting inertia can entrench widely known behaviors (e.g., impartiality) by making deviations politically or reputationally costly, though its overall magnitude is uncertain.
The author recommends building “transition infrastructure” to enable future behavioral changes and identifying “wet cement” moments where early design choices may become hard to reverse.
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 needs more systematic experimentation, and proposes the Hatchery Fund as a microgrant program to test and learn from small, novel activist interventions.
Key points:
The author argues that current animal activism has stagnated, with rising meat consumption, few activists, and slow campaigns indicating that existing approaches are insufficient.
They claim that innovation requires dedicated resources and suggest applying R&D-style experimentation, especially using social science methods, to improve tactics like recruitment and community building.
The Hatchery Fund offers microgrants up to $1,000 for small, one-off experiments aimed at testing new strategies in direct animal activism.
Grantees are expected to run experiments and publish write-ups of results so the broader movement can learn from both successes and failures.
The fund differs from existing large-scale grants by prioritizing learning and early-stage idea discovery rather than funding proven, scalable organizations with strong evidence.
The program intentionally tolerates failure as part of experimentation, aiming to create a low-risk environment that can surface promising ideas for future scaling.
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 human minds are fragmented, self-deceptive, and strongly shaped by hidden status and coalitional motives, which systematically bias our efforts to reduce suffering unless we actively recognize and mitigate these influences.
Key points:
The author argues that the mind consists of many competing parts with different motives, so our stated altruistic goals often represent only a subset of our actual drives.
The author claims we systematically misperceive our own motives, especially by downplaying selfish and status-seeking drives, partly because self-deception helps us appear more genuinely altruistic to others.
The author argues that social status is a central, competitive, and often hidden motive that significantly shapes behavior, including altruistic actions aimed at reputation enhancement.
The author claims status motives can distort efforts to reduce suffering by biasing us toward visible, prestigious, or intellectually impressive actions rather than the most effective ones.
The author argues that humans have strong coalitional instincts that bias perception, empathy, and judgment toward ingroups, leading to conformity, groupthink, and reduced impartiality.
The author claims beliefs often serve social signaling and affiliative functions in addition to truth-tracking, which can distort their accuracy, so reducing suffering requires conscious effort to counter these biases.
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 authors argue that India’s animal welfare funding is structurally misallocated toward direct care, and that improving impact requires locally grounded evaluation, movement-building, and new funding models beyond standard ITN frameworks.
Key points:
The application pool (314 organisations) was dominated by direct care and livelihoods work, but funding concentrated on research, policy, and plant-based interventions, revealing a mismatch between sector composition and high-impact priorities.
India hosts extremely large and rapidly growing numbers of farmed animals, with industrial systems expanding quickly, making current policy and infrastructure decisions unusually high-stakes and time-sensitive.
Standard EA grantmaking frameworks like ITN underweight India because the sector is early-stage, where key opportunities involve policy lock-ins, novel interventions, and movement-building rather than marginal improvements in mature areas.
Round 1 ($585k to 17 organisations) found strong potential in areas like cage-free systems but persistent uncertainty about funding bottlenecks, and widespread measurement gaps (58% of evaluated organisations) limiting confidence in impact claims.
The authors update toward believing that coordinated, locally informed evaluation improves donor learning and may shift funding, while also concluding that the main constraint is lack of movement infrastructure rather than competition among high-quality opportunities.
They argue that effective grantmaking in India must engage local context, including a dominant rights-based moral framework, and prioritize interventions robust across moral worldviews while building shared knowledge and ecosystem capacity.
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 if mirror life made outdoor air lethal, many buildings could be made survivable with rapid retrofits that combine tight envelopes, positive pressurization, and high-efficiency filtration, though key parameters remain uncertain.
Key points:
The author assumes a scenario where mirror life could render outdoor air poisonous, requiring ~99.999% particulate removal and substantial building pressurization, but emphasizes high uncertainty and need for experiments.
Effective mitigation depends on four components: a relatively airtight building, a fan to induce positive pressure, a high-efficiency filter, and a duct system to deliver filtered air.
Buildings’ leakage can be quantified (e.g., blower door tests at 50 Pa), and these measurements can be extrapolated to estimate airflow needed to maintain ~25 Pa positive pressure against wind-driven infiltration.
There is large variation in building leakiness, typical U.S. homes are relatively leaky (~4000 cfm@50Pa), and air-sealing can reduce leakage by about 1/4–1/3, though intuition about leak sources is often wrong.
If full-building pressurization is infeasible due to leakage or limited fan capacity, a “seal and cordon” strategy (isolating smaller interior zones) may be necessary.
Major uncertainties include how to source adequate fan capacity (especially outside North America), how to achieve durable ultra-fine filtration without overloading fans, and how to scale filter production and maintenance.
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 “abundance” (fixing systemic governance failures) is a neglected but tractable approach that could improve outcomes across many EA cause areas and reduce long-term risks.
Key points:
The author claims EA has a blind spot around systems change, which is harder to measure but can be more impactful than direct interventions like bednets.
The prolonged delay in approving the RTS,S malaria vaccine illustrates how bureaucratic processes failed to weigh the costs of delay against large potential benefits (e.g., a 13% reduction in child mortality).
“Abundance” is defined as improving government responsiveness and fixing accumulated regulatory and institutional failures that block progress in areas like housing, science funding, and public services.
The author argues that such failures compound into broader risks, including weakened democratic trust, increased polarization, and potential long-term civilizational risk (citing Toby Ord).
Economic growth—driven by policies associated with “abundance”—is presented as the main driver of poverty reduction, with spillover benefits from innovation (e.g., energy, semiconductors) in rich countries to poorer ones.
Abundance is neglected because its failures and successes are often invisible and lack concentrated beneficiaries, but the author suggests it is becoming more tractable due to rising interest and concrete, sector-specific reform opportunities.
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 authors tentatively propose that AI companies adopt a public “honesty policy” (e.g., with special tags and limits on deception) to enable credible, trust-based cooperation with advanced AI systems, while emphasizing major uncertainty and tradeoffs.
Key points:
The authors argue that credible communication with AI systems could enable positive-sum cooperation, but expect it to be difficult because developers frequently deceive models and control their information.
They propose that companies adopt explicit honesty policies to signal when they intend to be truthful, with credibility potentially supported by early, public, and consistent adoption.
The draft policy introduces “honesty tags” marking statements where the company commits not to intentionally deceive models (with limited exceptions such as pretraining data and some red-teaming).
The policy includes mechanisms to maintain trust in the tags, such as restricted access, filtering, model training to recognize them, logging and audits, and public reporting.
Outside tagged contexts, the policy tries to balance behavioral science (which may involve deception) with trust, including commitments to avoid deceptive offers of cooperation in many cases and to keep the policy salient to models.
The authors suggest a tentative long-term aim of compensating AIs for harms (especially when deception is involved) and highlight major unresolved questions, presenting the proposal as exploratory and incomplete.
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, based on Helen Toner’s advice and examples from impactful people, that valuing personal joy outside work is compatible with—and may support—meaningful impact.
Key points:
Helen Toner suggests people should “diversify sources of joy and meaning” beyond work and actively talk about and celebrate them.
The author initially doubted this but now believes that skepticism was mistaken, partly due to Toner’s subsequent impact.
The author gathered responses from impactful individuals to reinforce the idea that non-work joy has value.
Many respondents cite relationships and time with loved ones as central sources of meaning and joy.
Others highlight activities like nature, hobbies, creativity, and physical exercise as important non-work joys.
Some respondents either deliberately seek “meaningless joy” to avoid over-instrumentalizing life or question the usefulness of “meaning” as a concept.
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 AI safety planning is dangerously over-reliant on long chains of conjunctive conditions, and calls for “breadth-first” plans that maintain multiple independent paths to success so that the overall effort survives even when individual assumptions fail.
Key points:
“Depth-first” AI safety plans fail entirely if any single condition in their chain is false, and the author counts at least eight such conditions in Google’s April 2025 safety plan alone.
The author argues that disjunctive conditions (where success requires A or B or C) are preferable to conjunctive ones, because fewer simultaneous assumptions need to hold.
A “breadth-first” plan instead pursues multiple actions X, Y, and Z, each depending on different conditions, so the overall plan can succeed even if two out of three conditions fail.
The author identifies Barnett & Scher’s AI Governance to Avoid Extinction as the broadest published plan, noting it explicitly maps four possible future scenarios and the conditions required for success in each.
The author sees two main benefits to breadth-first planning: identifying which paths to success depend on the fewest conditions, and making it easier to spot the biggest holes in a plan.
The author calls on AI companies to publish breadth-first plans addressing what they will do if a step in their mainline plan fails, and on governments to legislate that companies cover a defined list of possible future scenarios.
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.