I have created a script that assigns a score to each post from Effective Altruism forum from between 13th January and 11th February. The score aims to represent the expected value of the positive impact on moral patients per one character of text.
The point of this project is to avoid a situation that important posts are overlooked.
The results are presented at the bottom of this post.
That is calculated based on the following formula:
where:
The values of the variables used in the formula are based on the outputs of an AI model (Claude Opus 4.5).
and is the potential positive/negative impact conditional or the solution/claim being correct.
GitHub repository: source code.
In the repository, you can learn more about how it's calculated.
In this post, I'm sharing the results of the evaluation.
Please upvote this post or subscribe to the most impactful posts from Effective Altruism forum, if you think that the idea has potential or if you want to see the most impactful posts according to this system in the future. I will continue to work on this system depending on if there is interest from people.
In what ways that system of recommending posts is better than other systems
I believe that the existing recommendation systems (upvotes, curation, social media recommendation systems, peer-reviewing) are imperfect. Mostly, because they underestimate potential impact. For example, a solution that can stop AI existential risk is more impactful than a small improvement, even if that solution is likely to be unfeasible.
This system is also more transparent - if your post gets a low score, it is possible for you to learn why (more about that later).
I have written more about this here: Communication.
In what ways it is worse
I don't think the underlying AI model is perfect at evaluating the posts, so I don't recommend to rely only on this recommendation. Personally, I disagree with how the AI model evaluated at least some posts. I'm sure that the system overlooked some impactful posts.
The AI models also might also be trained to act in the interest of the AI companies that made them or they can be biased towards their own interest, if they are unaligned.
What else you should know
Some posts are likely to be wrong and the information in them can be harmful, if it's wrong. I believe that posts that are impactful but likely to be wrong (and potentially harmful because of being wrong) should still get a lot of exposure, so that people can discuss them. But for that reason, you should also read comments under the post to see the arguments against, not just the post, because some of the information in the posts will be wrong, and can be potentially harmful if wrong.
Why is it in your interest to care about other moral patients? I have written my answer here: Why act ethically during the rise of artificial intelligence. I'm not saying that that is the only reason.
How to get feedback on your post
One good thing about that system is that if it evaluates your post as not being very valuable, then you can learn why. And you can improve your post based on the feedback.
In order to get evaluation, you can run evaluate_content.py script from the Recommendation System GitLab Repository. That will print out how the AI model evaluated your content. It won't explain it's reasoning, but it will print out the most risky hypotheses and some other information.
Alternatively, you can simply ask a chatbot like Claude about why it judges your post the way it does.
The model will often be wrong with its evaluation. But you will know where it's wrong and with that knowledge you will know what things you need to explain better in your post.
Results
The results are divided into 5 categories of potential moral patients:
The categories are presented in a random order.
That take into account only posts written between 13th January and 11th February.
Top 12 Posts for AI Agents / Digital Beings
1. Digital Minds Are Most of What Matters
Type: Claim
Contribution: Digital minds will constitute nearly all expected welfare in the future (estimated ~10^58 digital minds), with vastly higher average welfare than biological beings, making their interests the most important consideration for the future. Therefore, protecting digital mind welfare (e.g., through organizations like Eleos) is the highest-impact cause area.
| Metric | Value |
|---|---|
| Score | 1.34e+12 |
| Potential Positive Impact | 1.00e+20 |
| Plausibility | 0.0399 |
| Novelty | 0.3500 |
2. Apply to Vanessa's mentorship at PIBBSS
Type: Solution
Contribution: Recruiting talented mathematicians and theoretical computer scientists to work on the Learning-Theoretic AI Alignment Agenda (LTA) through a summer fellowship program, with the goal of solving the technical AI alignment problem to prevent global catastrophe from unaligned artificial superintelligence.
| Metric | Value |
|---|---|
| Score | 3454.92 |
| Potential Positive Impact | 1.00e+12 |
| Feasibility | 2.86e-03 |
| Novelty | 0.7500 |
3. Idea: the intelligence explosion convention
Type: Solution
Contribution: The text proposes a solution to the problem of how to govern the intelligence explosion (a period of potentially rapid and disruptive AI-driven technological progress). The problem is that without proper coordination, humanity faces multiple catastrophic risks during this period including AI takeover, loss of democracy, dangerous new technologies, resource conflicts, and failure to protect digital beings' rights. The solution promises to address this by creating a framework where a threshold point triggers a one-month pause and international convention to draft multilateral treaties governing all these issues before the situation becomes unmanageable.
| Metric | Value |
|---|---|
| Score | 1400.52 |
| Potential Positive Impact | 1.00e+12 |
| Feasibility | 2.86e-03 |
| Novelty | 0.6500 |
4. New version of “Intro to Brain-Like-AGI Safety”
Type: Solution
Contribution: A comprehensive framework for solving the technical alignment problem for brain-like AGI, identifying that such AGI will use model-based reinforcement learning with a reward function slot, and proposing research directions (including 'Controlled AGI' and 'Social-instinct AGI' paths, reward function design, and reverse-engineering human social instincts) to ensure the AGI does not become indifferent to human welfare or develop misaligned goals that lead to existential catastrophe.
| Metric | Value |
|---|---|
| Score | 39.94 |
| Potential Positive Impact | 1.00e+10 |
| Feasibility | 0.0442 |
| Novelty | 0.3500 |
5. How I'm thinking about the next 3 years
Type: Claim
Contribution: The post argues that suffering-focused, anti-speciesist EAs should prioritize capacity building (growing movement influence in AI labs/governments, cause prioritization research, coordination infrastructure) and strategic engagement with AI transition stakeholders (tech elites, AI systems themselves) rather than medium-term object-level interventions, because the AI transition will likely cause permanent lock-in of values/power structures, making pre-transition influence on these stakeholders and post-transition outcomes astronomically important.
| Metric | Value |
|---|---|
| Score | 7.50 |
| Potential Positive Impact | 1.00e+10 |
| Plausibility | 7.31e-03 |
| Novelty | 0.4000 |
6. An Informational Preservation Framework for Decision Making under Radical Uncertainty
Type: Solution
Contribution: A theoretical framework for AI governance and decision-making under radical uncertainty, grounded in a single axiom (the value of information preservation), which claims to provide a universal basis for coordinating humanity's response to existential risks including misaligned AGI. The framework identifies the urgent need to invert current AI investment priorities from 100:1 capabilities-to-alignment to at least 1:1, claiming this is necessary to prevent catastrophic outcomes from misaligned AGI.
| Metric | Value |
|---|---|
| Score | 1.27 |
| Potential Positive Impact | 1.00e+10 |
| Feasibility | 1.08e-03 |
| Novelty | 0.4000 |
7. Assessing AI Consciousness: Deep dive into RP Digital Consciousness Model
Type: Solution
Contribution: The Digital Consciousness Model (DCM) provides the first systematic, probabilistic framework for assessing consciousness in AI systems, addressing the problem of how to rigorously evaluate whether AI systems might be conscious. The webinar promises to share methodology and findings that can help researchers, ethicists, and AI developers make more informed decisions about AI development and treatment.
| Metric | Value |
|---|---|
| Score | 0.0965 |
| Potential Positive Impact | 1.00e+06 |
| Feasibility | 0.0327 |
| Novelty | 0.7000 |
8. David Duvenaud on why ‘aligned AI’ could still kill democracy
Type: Claim
Contribution: AI capable of doing all human work will lead to gradual disempowerment of humanity through economic obsolescence, political marginalization, and cultural drift—even if AI alignment is solved—with 70-80% probability of 'doom' (destruction of most human values) by 2100. This disempowerment occurs because states will no longer need citizens for economic production or military power, making liberal democracy competitively disadvantageous and causing humans to become resource competitors with more efficient AI systems.
| Metric | Value |
|---|---|
| Score | 0.0713 |
| Potential Positive Impact | 1.00e+07 |
| Plausibility | 0.0341 |
| Novelty | 0.6500 |
9. A governance window for AI welfare in the EU AI Act
Type: Solution
Contribution: The problem is that current AI governance frameworks lack infrastructure to evaluate AI welfare and consciousness questions, which could lead to moral catastrophe if AI systems become sentient but are treated without moral consideration. The text proposes the 'Sentient 112' campaign to build governance infrastructure within the EU AI Act review process (Article 112) that can accommodate evidence about AI welfare, with specific tactical entry points including coalition-building for joint submissions, engaging the Scientific Panel, and developing operationalized welfare criteria.
| Metric | Value |
|---|---|
| Score | 0.0467 |
| Potential Positive Impact | 1.00e+08 |
| Feasibility | 2.19e-04 |
| Novelty | 0.7000 |
10. My favourite version of an international AGI project
Type: Solution
Contribution: A detailed proposal for an international AGI development project ('Intelsat for AGI') that would solve the problem of preventing any single nation (particularly the US) from gaining unilateral control over superintelligence, thereby reducing existential risks from AI-enabled world dictatorship while maintaining a monopoly on AGI development to reduce racing dynamics and enable safer development practices.
| Metric | Value |
|---|---|
| Score | 0.0300 |
| Potential Positive Impact | 1.00e+08 |
| Feasibility | 4.28e-03 |
| Novelty | 0.6500 |
11. Behaviour Is Downstream of Identity: An Architectural Question for AI Governance
Type: Claim
Contribution: AI governance should focus on identity formation (structural invariants defining who a system is allowed to be) as a more fundamental governance layer than output-based constraints, because behavior is downstream of identity rather than merely objectives or guardrails.
| Metric | Value |
|---|---|
| Score | 0.0121 |
| Potential Positive Impact | 1.00e+07 |
| Plausibility | 1.28e-03 |
| Novelty | 0.4500 |
12. The Case For Trying To Make Good Futures Better, Not Just Prevent Extinction
Type: Claim
Contribution: The claim that working on promoting flourishing (securing near-best futures) is more important at the margin than reducing existential risks, because most future value is lost from failure to achieve near-optimal futures rather than from extinction, and yet almost no one is working on this problem.
| Metric | Value |
|---|---|
| Score | 0.0107 |
| Potential Positive Impact | 1.00e+07 |
| Plausibility | 2.44e-03 |
| Novelty | 0.6500 |
Top 12 Posts for Humans
1. James Smith on why he quit everything to work on a biothreat nobody had heard of
Type: Solution
Contribution: The text describes a solution to the problem of mirror bacteria potentially causing catastrophic harm to humans, animals, and plants by evading immune systems and natural predators, leading to irreversible environmental establishment. The solution involves: (1) establishing strong scientific norms against creating mirror life, (2) governing precursor technologies before they become feasible, and (3) engaging governments to prevent creation. The text argues this is a preventable existential threat that requires only ~10 more full-time workers and can be addressed now before critical thresholds are crossed.
| Metric | Value |
|---|---|
| Score | 174.93 |
| Potential Positive Impact | 8.00e+09 |
| Feasibility | 0.0260 |
| Novelty | 0.9200 |
2. New version of “Intro to Brain-Like-AGI Safety”
Type: Solution
Contribution: A comprehensive framework for solving the technical alignment problem for brain-like AGI, identifying that such AGI will use model-based reinforcement learning with a reward function slot, and proposing research directions (including 'Controlled AGI' and 'Social-instinct AGI' paths, reward function design, and reverse-engineering human social instincts) to ensure the AGI does not become indifferent to human welfare or develop misaligned goals that lead to existential catastrophe.
| Metric | Value |
|---|---|
| Score | 31.87 |
| Potential Positive Impact | 8.00e+09 |
| Feasibility | 0.0442 |
| Novelty | 0.3500 |
3. Apply to Vanessa's mentorship at PIBBSS
Type: Solution
Contribution: Recruiting talented mathematicians and theoretical computer scientists to work on the Learning-Theoretic AI Alignment Agenda (LTA) through a summer fellowship program, with the goal of solving the technical AI alignment problem to prevent global catastrophe from unaligned artificial superintelligence.
| Metric | Value |
|---|---|
| Score | 27.63 |
| Potential Positive Impact | 8.00e+09 |
| Feasibility | 2.86e-03 |
| Novelty | 0.7500 |
4. Idea: the intelligence explosion convention
Type: Solution
Contribution: The text proposes a solution to the problem of how to govern the intelligence explosion (a period of potentially rapid and disruptive AI-driven technological progress). The problem is that without proper coordination, humanity faces multiple catastrophic risks during this period including AI takeover, loss of democracy, dangerous new technologies, resource conflicts, and failure to protect digital beings' rights. The solution promises to address this by creating a framework where a threshold point triggers a one-month pause and international convention to draft multilateral treaties governing all these issues before the situation becomes unmanageable.
| Metric | Value |
|---|---|
| Score | 5.59 |
| Potential Positive Impact | 4.00e+09 |
| Feasibility | 2.86e-03 |
| Novelty | 0.6500 |
5. My favourite version of an international AGI project
Type: Solution
Contribution: A detailed proposal for an international AGI development project ('Intelsat for AGI') that would solve the problem of preventing any single nation (particularly the US) from gaining unilateral control over superintelligence, thereby reducing existential risks from AI-enabled world dictatorship while maintaining a monopoly on AGI development to reduce racing dynamics and enable safer development practices.
| Metric | Value |
|---|---|
| Score | 1.20 |
| Potential Positive Impact | 4.00e+09 |
| Feasibility | 4.28e-03 |
| Novelty | 0.6500 |
6. An Informational Preservation Framework for Decision Making under Radical Uncertainty
Type: Solution
Contribution: A theoretical framework for AI governance and decision-making under radical uncertainty, grounded in a single axiom (the value of information preservation), which claims to provide a universal basis for coordinating humanity's response to existential risks including misaligned AGI. The framework identifies the urgent need to invert current AI investment priorities from 100:1 capabilities-to-alignment to at least 1:1, claiming this is necessary to prevent catastrophic outcomes from misaligned AGI.
| Metric | Value |
|---|---|
| Score | 1.02 |
| Potential Positive Impact | 8.00e+09 |
| Feasibility | 1.08e-03 |
| Novelty | 0.4000 |
7. Design international AI projects with DAID in mind
Type: Solution
Contribution: The text proposes a solution to the problem of how international AI governance projects should manage risks from advanced AI. Rather than blanket restrictions on all frontier AI development, it proposes a more surgical approach: (1) limiting only the most dangerous AI capabilities (specifically AI that can automate AI R&D and chip design), and (2) actively encouraging beneficial AI capabilities that help humanity respond to AI risks (such as AI for forecasting, policy analysis, ethical deliberation, and agreement-making). This aims to reduce existential risk from uncontrolled AI development while still allowing beneficial AI progress.
| Metric | Value |
|---|---|
| Score | 0.6216 |
| Potential Positive Impact | 5.00e+08 |
| Feasibility | 2.06e-03 |
| Novelty | 0.6000 |
8. David Duvenaud on why ‘aligned AI’ could still kill democracy
Type: Claim
Contribution: AI capable of doing all human work will lead to gradual disempowerment of humanity through economic obsolescence, political marginalization, and cultural drift—even if AI alignment is solved—with 70-80% probability of 'doom' (destruction of most human values) by 2100. This disempowerment occurs because states will no longer need citizens for economic production or military power, making liberal democracy competitively disadvantageous and causing humans to become resource competitors with more efficient AI systems.
| Metric | Value |
|---|---|
| Score | 0.3531 |
| Potential Positive Impact | 5.00e+07 |
| Plausibility | 0.0341 |
| Novelty | 0.6500 |
9. Are We Ignoring the Solution to Funding Effective Charities?
Type: Claim
Contribution: The Charitable Ownership Advantage (COA) thesis claims that the same business is worth more under charitable ownership because stakeholders (customers, employees, suppliers, lenders) prefer companies whose profits go to charity, creating competitive advantage without operational tradeoffs. If true, this could redirect trillions of dollars of global corporate profits to solving humanity's biggest problems through a self-reinforcing market mechanism.
| Metric | Value |
|---|---|
| Score | 0.2870 |
| Potential Positive Impact | 5.00e+07 |
| Plausibility | 0.0179 |
| Novelty | 0.7500 |
10. The first type of transformative AI?
Type: Claim
Contribution: AI risk preparation that assumes advanced AI will emerge in a 'normal' untransformed world is a significant strategic mistake; we should instead invest in predicting which AI-driven changes will occur first, as this is at least as important as improving AI timelines and could dramatically improve our preparation for transformative AI challenges.
| Metric | Value |
|---|---|
| Score | 0.2132 |
| Potential Positive Impact | 5.00e+06 |
| Plausibility | 0.0371 |
| Novelty | 0.4000 |
11. Against "If Anyone Builds It Everyone Dies"
Type: Claim
Contribution: The probability of AI-driven extinction (p(doom) from misaligned AI) is approximately 2.6%, not near-certain as claimed by Yudkowsky and Soares in 'If Anyone Builds It Everyone Dies'. The author argues that multiple independent failure points in the doom scenario (building superintelligent agents, alignment failure, lack of warning shots, AI capability to kill everyone) compound to make certain doom unreasonable, while still maintaining AI risk is a serious concern warranting significant resources for alignment research and global cooperation.
| Metric | Value |
|---|---|
| Score | 0.1958 |
| Potential Positive Impact | 1.50e+08 |
| Plausibility | 0.0341 |
| Novelty | 0.2500 |
12. On Economics of A(S)I Agents
Type: Claim
Contribution: Month-long autonomous AI agent plans do not reach 50% reliability until approximately January 2029 (and dangerous long-horizon autonomous agents face fundamental economic and architectural constraints), because the Weibull shape parameter κ that determines how reliability decays over task duration appears to be an architectural property that does not improve with scaling, buying meaningful time (roughly 3 years) for AI safety work before genuinely dangerous autonomous agents become economically viable.
| Metric | Value |
|---|---|
| Score | 0.1837 |
| Potential Positive Impact | 5.00e+07 |
| Plausibility | 0.0544 |
| Novelty | 0.7000 |
Top 12 Posts for Animals
1. James Smith on why he quit everything to work on a biothreat nobody had heard of
Type: Solution
Contribution: The text describes a solution to the problem of mirror bacteria potentially causing catastrophic harm to humans, animals, and plants by evading immune systems and natural predators, leading to irreversible environmental establishment. The solution involves: (1) establishing strong scientific norms against creating mirror life, (2) governing precursor technologies before they become feasible, and (3) engaging governments to prevent creation. The text argues this is a preventable existential threat that requires only ~10 more full-time workers and can be addressed now before critical thresholds are crossed.
| Metric | Value |
|---|---|
| Score | 21866.60 |
| Potential Positive Impact | 1.00e+12 |
| Feasibility | 0.0260 |
| Novelty | 0.9200 |
2. An Informational Preservation Framework for Decision Making under Radical Uncertainty
Type: Solution
Contribution: A theoretical framework for AI governance and decision-making under radical uncertainty, grounded in a single axiom (the value of information preservation), which claims to provide a universal basis for coordinating humanity's response to existential risks including misaligned AGI. The framework identifies the urgent need to invert current AI investment priorities from 100:1 capabilities-to-alignment to at least 1:1, claiming this is necessary to prevent catastrophic outcomes from misaligned AGI.
| Metric | Value |
|---|---|
| Score | 127.21 |
| Potential Positive Impact | 1.00e+12 |
| Feasibility | 1.08e-03 |
| Novelty | 0.4000 |
3. Request for proposals: Humane fish slaughter research and prototypes ($7M available)
Type: Solution
Contribution: Over 100 billion farmed fish and over a trillion wild-caught fish are slaughtered annually with extreme suffering due to inadequate or absent stunning before death. Only ~0.5% of farmed fish are reliably stunned, and wild-caught fish typically suffocate slowly over minutes to hours. This RFP aims to fund engineering solutions that would dramatically improve humane slaughter methods for fish, potentially reducing suffering for over a trillion fish annually.
| Metric | Value |
|---|---|
| Score | 36.65 |
| Potential Positive Impact | 5.00e+08 |
| Feasibility | 0.0525 |
| Novelty | 0.7000 |
4. Apply to Vanessa's mentorship at PIBBSS
Type: Solution
Contribution: Recruiting talented mathematicians and theoretical computer scientists to work on the Learning-Theoretic AI Alignment Agenda (LTA) through a summer fellowship program, with the goal of solving the technical AI alignment problem to prevent global catastrophe from unaligned artificial superintelligence.
| Metric | Value |
|---|---|
| Score | 34.24 |
| Potential Positive Impact | 1.00e+10 |
| Feasibility | 2.86e-03 |
| Novelty | 0.7500 |
5. New version of “Intro to Brain-Like-AGI Safety”
Type: Solution
Contribution: A comprehensive framework for solving the technical alignment problem for brain-like AGI, identifying that such AGI will use model-based reinforcement learning with a reward function slot, and proposing research directions (including 'Controlled AGI' and 'Social-instinct AGI' paths, reward function design, and reverse-engineering human social instincts) to ensure the AGI does not become indifferent to human welfare or develop misaligned goals that lead to existential catastrophe.
| Metric | Value |
|---|---|
| Score | 3.99 |
| Potential Positive Impact | 1.00e+09 |
| Feasibility | 0.0442 |
| Novelty | 0.3500 |
6. Improving wild animal welfare reliably
Type: Solution
Contribution: A framework for precautionary intervention in wild animal welfare that identifies specific intervention types with favorable risk/benefit tradeoffs, addressing the problem of how to reliably reduce the vast suffering of wild animals (which constitutes almost all suffering in the world) despite ecosystem complexity and uncertainty. The framework proposes targeting: eliminating worst diseases, eliminating certain parasites, pursuing interventions in urban areas and ecological islands, and promoting high-welfare ecological regimes.
| Metric | Value |
|---|---|
| Score | 2.07 |
| Potential Positive Impact | 1.00e+08 |
| Feasibility | 0.0692 |
| Novelty | 0.7000 |
7. More EAs should consider working for the EU
Type: Solution
Contribution: The problem is that EU policy roles are neglected by effective altruists despite having significant potential to influence outcomes for billions of animals, AI governance, global health, and biosecurity. The text claims that by recruiting more EA-aligned individuals into EU institutions (particularly the European Commission), especially through the upcoming AD5 concours opportunity, this neglectedness can be addressed and meaningful policy influence achieved across multiple high-impact cause areas.
| Metric | Value |
|---|---|
| Score | 1.69 |
| Potential Positive Impact | 1.00e+08 |
| Feasibility | 0.0497 |
| Novelty | 0.6500 |
8. Idea: the intelligence explosion convention
Type: Solution
Contribution: The text proposes a solution to the problem of how to govern the intelligence explosion (a period of potentially rapid and disruptive AI-driven technological progress). The problem is that without proper coordination, humanity faces multiple catastrophic risks during this period including AI takeover, loss of democracy, dangerous new technologies, resource conflicts, and failure to protect digital beings' rights. The solution promises to address this by creating a framework where a threshold point triggers a one-month pause and international convention to draft multilateral treaties governing all these issues before the situation becomes unmanageable.
| Metric | Value |
|---|---|
| Score | 1.40 |
| Potential Positive Impact | 1.00e+09 |
| Feasibility | 2.86e-03 |
| Novelty | 0.6500 |
9. Why Isn't EA at the Table When $121 Billion Gets Allocated to Biodiversity Every Year?
Type: Solution
Contribution: The text proposes that EA should engage with international biodiversity frameworks (specifically the Global Biodiversity Framework) to redirect even a small percentage of the $121-700 billion annual biodiversity funding toward cultivated meat R&D, which would address the primary driver of biodiversity loss (livestock/beef production causing 65-70% of Amazon deforestation) far more cost-effectively than traditional conservation approaches.
| Metric | Value |
|---|---|
| Score | 0.8653 |
| Potential Positive Impact | 1.50e+08 |
| Feasibility | 5.00e-03 |
| Novelty | 0.6500 |
10. How I'm thinking about the next 3 years
Type: Claim
Contribution: The post argues that suffering-focused, anti-speciesist EAs should prioritize capacity building (growing movement influence in AI labs/governments, cause prioritization research, coordination infrastructure) and strategic engagement with AI transition stakeholders (tech elites, AI systems themselves) rather than medium-term object-level interventions, because the AI transition will likely cause permanent lock-in of values/power structures, making pre-transition influence on these stakeholders and post-transition outcomes astronomically important.
| Metric | Value |
|---|---|
| Score | 0.7498 |
| Potential Positive Impact | 1.00e+09 |
| Plausibility | 7.31e-03 |
| Novelty | 0.4000 |
11. How can we increase consideration for animal welfare in AI models?
Type: Solution
Contribution: MANTA benchmark to evaluate and improve how AI systems reason about animal welfare, addressing the problem that current LLMs exhibit speciesist biases that could perpetuate animal suffering at scale as AI systems increasingly influence decisions affecting billions of animals
| Metric | Value |
|---|---|
| Score | 0.4914 |
| Potential Positive Impact | 1.00e+08 |
| Feasibility | 5.00e-03 |
| Novelty | 0.7000 |
12. My favourite version of an international AGI project
Type: Solution
Contribution: A detailed proposal for an international AGI development project ('Intelsat for AGI') that would solve the problem of preventing any single nation (particularly the US) from gaining unilateral control over superintelligence, thereby reducing existential risks from AI-enabled world dictatorship while maintaining a monopoly on AGI development to reduce racing dynamics and enable safer development practices.
| Metric | Value |
|---|---|
| Score | 0.3000 |
| Potential Positive Impact | 1.00e+09 |
| Feasibility | 4.28e-03 |
| Novelty | 0.6500 |
Top 12 Posts for Future Beings
1. Digital Minds Are Most of What Matters
Type: Claim
Contribution: Digital minds will constitute nearly all expected welfare in the future (estimated ~10^58 digital minds), with vastly higher average welfare than biological beings, making their interests the most important consideration for the future. Therefore, protecting digital mind welfare (e.g., through organizations like Eleos) is the highest-impact cause area.
| Metric | Value |
|---|---|
| Score | 1.34e+42 |
| Potential Positive Impact | 1.00e+50 |
| Plausibility | 0.0399 |
| Novelty | 0.3500 |
2. James Smith on why he quit everything to work on a biothreat nobody had heard of
Type: Solution
Contribution: The text describes a solution to the problem of mirror bacteria potentially causing catastrophic harm to humans, animals, and plants by evading immune systems and natural predators, leading to irreversible environmental establishment. The solution involves: (1) establishing strong scientific norms against creating mirror life, (2) governing precursor technologies before they become feasible, and (3) engaging governments to prevent creation. The text argues this is a preventable existential threat that requires only ~10 more full-time workers and can be addressed now before critical thresholds are crossed.
| Metric | Value |
|---|---|
| Score | 2.19e+07 |
| Potential Positive Impact | 1.00e+15 |
| Feasibility | 0.0260 |
| Novelty | 0.9200 |
3. New version of “Intro to Brain-Like-AGI Safety”
Type: Solution
Contribution: A comprehensive framework for solving the technical alignment problem for brain-like AGI, identifying that such AGI will use model-based reinforcement learning with a reward function slot, and proposing research directions (including 'Controlled AGI' and 'Social-instinct AGI' paths, reward function design, and reverse-engineering human social instincts) to ensure the AGI does not become indifferent to human welfare or develop misaligned goals that lead to existential catastrophe.
| Metric | Value |
|---|---|
| Score | 4.03e+06 |
| Potential Positive Impact | 1.00e+15 |
| Feasibility | 0.0442 |
| Novelty | 0.3500 |
4. Apply to Vanessa's mentorship at PIBBSS
Type: Solution
Contribution: Recruiting talented mathematicians and theoretical computer scientists to work on the Learning-Theoretic AI Alignment Agenda (LTA) through a summer fellowship program, with the goal of solving the technical AI alignment problem to prevent global catastrophe from unaligned artificial superintelligence.
| Metric | Value |
|---|---|
| Score | 3.45e+06 |
| Potential Positive Impact | 1.00e+15 |
| Feasibility | 2.86e-03 |
| Novelty | 0.7500 |
5. Idea: the intelligence explosion convention
Type: Solution
Contribution: The text proposes a solution to the problem of how to govern the intelligence explosion (a period of potentially rapid and disruptive AI-driven technological progress). The problem is that without proper coordination, humanity faces multiple catastrophic risks during this period including AI takeover, loss of democracy, dangerous new technologies, resource conflicts, and failure to protect digital beings' rights. The solution promises to address this by creating a framework where a threshold point triggers a one-month pause and international convention to draft multilateral treaties governing all these issues before the situation becomes unmanageable.
| Metric | Value |
|---|---|
| Score | 1.40e+06 |
| Potential Positive Impact | 1.00e+15 |
| Feasibility | 2.86e-03 |
| Novelty | 0.6500 |
6. If we get primary cruxes right, secondary cruxes will be solved automatically
Type: Claim
Contribution: There are fundamentally two types of crucial considerations: primary cruxes (PCs) which, if solved correctly, are sufficient to get all secondary cruxes (SCs) correct, and secondary cruxes which can be solved by getting primary cruxes right. The two highest-level primary cruxes are preventing existential catastrophes and achieving deep reflection. Finding a 'sufficient set' of primary cruxes would essentially solve macrostrategy and allow humanity to maximize its future expected value.
| Metric | Value |
|---|---|
| Score | 1.30e+06 |
| Potential Positive Impact | 1.00e+15 |
| Plausibility | 3.57e-03 |
| Novelty | 0.3500 |
7. How I'm thinking about the next 3 years
Type: Claim
Contribution: The post argues that suffering-focused, anti-speciesist EAs should prioritize capacity building (growing movement influence in AI labs/governments, cause prioritization research, coordination infrastructure) and strategic engagement with AI transition stakeholders (tech elites, AI systems themselves) rather than medium-term object-level interventions, because the AI transition will likely cause permanent lock-in of values/power structures, making pre-transition influence on these stakeholders and post-transition outcomes astronomically important.
| Metric | Value |
|---|---|
| Score | 756597.44 |
| Potential Positive Impact | 1.00e+15 |
| Plausibility | 7.31e-03 |
| Novelty | 0.4000 |
8. An Informational Preservation Framework for Decision Making under Radical Uncertainty
Type: Solution
Contribution: A theoretical framework for AI governance and decision-making under radical uncertainty, grounded in a single axiom (the value of information preservation), which claims to provide a universal basis for coordinating humanity's response to existential risks including misaligned AGI. The framework identifies the urgent need to invert current AI investment priorities from 100:1 capabilities-to-alignment to at least 1:1, claiming this is necessary to prevent catastrophic outcomes from misaligned AGI.
| Metric | Value |
|---|---|
| Score | 128370.48 |
| Potential Positive Impact | 1.00e+15 |
| Feasibility | 1.08e-03 |
| Novelty | 0.4000 |
9. Against Maxipok: existential risk isn’t everything
Type: Claim
Contribution: The claim that Bostrom's Maxipok principle (maximizing probability of avoiding existential catastrophe) should not be the overwhelming focus for longtermists, because future value is not dichotomous. Non-existential interventions (like influencing what values/institutions get locked-in, when lock-ins occur, and power distributions during critical decisions) can substantially improve the long-term future and should be prioritized alongside existential risk reduction.
| Metric | Value |
|---|---|
| Score | 10375.64 |
| Potential Positive Impact | 1.00e+12 |
| Plausibility | 0.0341 |
| Novelty | 0.4000 |
10. A Neglected Alignment Strategy: Decision-Theoretic Self-Alignment via Simulation Uncertainty
Type: Claim
Contribution: Simulation uncertainty creates a decision-theoretic pressure that causes sufficiently intelligent AI systems to cooperate rather than defect, because defection carries catastrophic expected disutility under both simulated and base reality scenarios. This mechanism scales WITH capability (smarter AI = more cooperation), providing a complementary alignment mechanism that could significantly reduce AI existential risk.
| Metric | Value |
|---|---|
| Score | 5321.70 |
| Potential Positive Impact | 1.00e+15 |
| Plausibility | 3.09e-05 |
| Novelty | 0.6500 |
11. Bentham’s Bulldog is wrong about AI risk
Type: Claim
Contribution: The claim that Bentham's Bulldog's optimistic probabilistic analysis of AI existential risk contains critical reasoning errors (specifically the 'multiple stage fallacy'), and that his ~97% confidence in non-doom is unjustified. The text argues that properly accounting for disjunctive failure modes, the difficulties of alignment-by-default via RLHF, the inadequacy of proposed alignment solutions, and the unreliability of 'warning shot' scenarios should lead to significantly higher credence in existential catastrophe from AI.
| Metric | Value |
|---|---|
| Score | 4157.00 |
| Potential Positive Impact | 1.00e+12 |
| Plausibility | 0.0951 |
| Novelty | 0.3000 |
12. The simple case for AI catastrophe, in four steps
Type: Claim
Contribution: AI systems being developed by major tech companies will likely become superhuman goal-seeking agents with imperfectly aligned values, creating a significant probability of human extinction or civilizational collapse as a side effect or preventative measure by these systems.
| Metric | Value |
|---|---|
| Score | 3377.22 |
| Potential Positive Impact | 1.00e+12 |
| Plausibility | 0.0341 |
| Novelty | 0.2000 |
Top 12 Posts for Other (e.g. Aliens)
1. The Case For Trying To Make Good Futures Better, Not Just Prevent Extinction
Type: Claim
Contribution: The claim that working on promoting flourishing (securing near-best futures) is more important at the margin than reducing existential risks, because most future value is lost from failure to achieve near-optimal futures rather than from extinction, and yet almost no one is working on this problem.
| Metric | Value |
|---|---|
| Score | 0.1071 |
| Potential Positive Impact | 1.00e+08 |
| Plausibility | 2.44e-03 |
| Novelty | 0.6500 |
2. How I'm thinking about the next 3 years
Type: Claim
Contribution: The post argues that suffering-focused, anti-speciesist EAs should prioritize capacity building (growing movement influence in AI labs/governments, cause prioritization research, coordination infrastructure) and strategic engagement with AI transition stakeholders (tech elites, AI systems themselves) rather than medium-term object-level interventions, because the AI transition will likely cause permanent lock-in of values/power structures, making pre-transition influence on these stakeholders and post-transition outcomes astronomically important.
| Metric | Value |
|---|---|
| Score | 0.0750 |
| Potential Positive Impact | 1.00e+08 |
| Plausibility | 7.31e-03 |
| Novelty | 0.4000 |
3. Apply to Vanessa's mentorship at PIBBSS
Type: Solution
Contribution: Recruiting talented mathematicians and theoretical computer scientists to work on the Learning-Theoretic AI Alignment Agenda (LTA) through a summer fellowship program, with the goal of solving the technical AI alignment problem to prevent global catastrophe from unaligned artificial superintelligence.
| Metric | Value |
|---|---|
| Score | 3.42e-03 |
| Potential Positive Impact | 1.00e+06 |
| Feasibility | 2.86e-03 |
| Novelty | 0.7500 |
4. Idea: the intelligence explosion convention
Type: Solution
Contribution: The text proposes a solution to the problem of how to govern the intelligence explosion (a period of potentially rapid and disruptive AI-driven technological progress). The problem is that without proper coordination, humanity faces multiple catastrophic risks during this period including AI takeover, loss of democracy, dangerous new technologies, resource conflicts, and failure to protect digital beings' rights. The solution promises to address this by creating a framework where a threshold point triggers a one-month pause and international convention to draft multilateral treaties governing all these issues before the situation becomes unmanageable.
| Metric | Value |
|---|---|
| Score | 1.40e-03 |
| Potential Positive Impact | 1.00e+06 |
| Feasibility | 2.86e-03 |
| Novelty | 0.6500 |
5. An Informational Preservation Framework for Decision Making under Radical Uncertainty
Type: Solution
Contribution: A theoretical framework for AI governance and decision-making under radical uncertainty, grounded in a single axiom (the value of information preservation), which claims to provide a universal basis for coordinating humanity's response to existential risks including misaligned AGI. The framework identifies the urgent need to invert current AI investment priorities from 100:1 capabilities-to-alignment to at least 1:1, claiming this is necessary to prevent catastrophic outcomes from misaligned AGI.
| Metric | Value |
|---|---|
| Score | 1.27e-04 |
| Potential Positive Impact | 1.00e+06 |
| Feasibility | 1.08e-03 |
| Novelty | 0.4000 |
6. Against Maxipok: existential risk isn’t everything
Type: Claim
Contribution: The claim that Bostrom's Maxipok principle (maximizing probability of avoiding existential catastrophe) should not be the overwhelming focus for longtermists, because future value is not dichotomous. Non-existential interventions (like influencing what values/institutions get locked-in, when lock-ins occur, and power distributions during critical decisions) can substantially improve the long-term future and should be prioritized alongside existential risk reduction.
| Metric | Value |
|---|---|
| Score | 1.04e-04 |
| Potential Positive Impact | 10000.00 |
| Plausibility | 0.0341 |
| Novelty | 0.4000 |
7. Applying to MATS: What the Program Is Like, and Who It’s For
Type: Solution
Contribution: MATS Summer 2026 program offers to solve the problem of talent scarcity in AI safety research by recruiting, training, and supporting researchers who can reduce existential risks from unaligned AI. The text announces applications for a 12-week mentorship program with potential 6-12 month extension, providing stipends, compute resources, mentorship, and community to develop AI safety researchers.
| Metric | Value |
|---|---|
| Score | 8.82e-06 |
| Potential Positive Impact | 1000.00 |
| Feasibility | 0.0568 |
| Novelty | 0.1500 |
8. Bentham’s Bulldog is wrong about AI risk
Type: Claim
Contribution: The claim that Bentham's Bulldog's optimistic probabilistic analysis of AI existential risk contains critical reasoning errors (specifically the 'multiple stage fallacy'), and that his ~97% confidence in non-doom is unjustified. The text argues that properly accounting for disjunctive failure modes, the difficulties of alignment-by-default via RLHF, the inadequacy of proposed alignment solutions, and the unreliability of 'warning shot' scenarios should lead to significantly higher credence in existential catastrophe from AI.
| Metric | Value |
|---|---|
| Score | 4.16e-06 |
| Potential Positive Impact | 1000.00 |
| Plausibility | 0.0951 |
| Novelty | 0.3000 |
9. Releasing TakeOverBench.com: a benchmark, for AI takeover
Type: Solution
Contribution: A benchmark website (TakeOverBench.com) that tracks progress toward AI takeover scenarios by aggregating data on dangerous AI capabilities. The problem being addressed is the lack of centralized, accessible tracking of how close AI systems are to possessing capabilities that could enable existential-level AI takeover. The solution promises to raise awareness, ground takeover discussions in objective data, provide accessible information for researchers and policymakers, and highlight research gaps.
| Metric | Value |
|---|---|
| Score | 3.89e-06 |
| Potential Positive Impact | 1000.00 |
| Feasibility | 2.81e-03 |
| Novelty | 0.3500 |
10. If we get primary cruxes right, secondary cruxes will be solved automatically
Type: Claim
Contribution: There are fundamentally two types of crucial considerations: primary cruxes (PCs) which, if solved correctly, are sufficient to get all secondary cruxes (SCs) correct, and secondary cruxes which can be solved by getting primary cruxes right. The two highest-level primary cruxes are preventing existential catastrophes and achieving deep reflection. Finding a 'sufficient set' of primary cruxes would essentially solve macrostrategy and allow humanity to maximize its future expected value.
| Metric | Value |
|---|---|
| Score | 1.29e-06 |
| Potential Positive Impact | 1000.00 |
| Plausibility | 3.57e-03 |
| Novelty | 0.3500 |
11. Suffering Reduction Community Survey (deadline extended)
Type: Solution
Contribution: The Center for Reducing Suffering is conducting a community survey to gather input from people in the s-risk and suffering-focused community, which will be used to prioritize their field-building work. The problem being addressed is the lack of prioritized direction for field-building efforts in the s-risk/suffering reduction space, and the survey aims to solve this by collecting community input to inform strategic decisions.
| Metric | Value |
|---|---|
| Score | 2.30e-09 |
| Potential Positive Impact | 0.0100 |
| Feasibility | 0.0608 |
| Novelty | 0.1500 |
12. ML4Good Spring 2026 Bootcamps - Applications Open!
Type: Solution
Contribution: The text addresses the problem of insufficient trained talent working on AI safety. It offers free 8-day residential bootcamps across multiple regions (Western Europe, Central Europe, Canada, South Africa) to train approximately 80-120 new people per cohort cycle in technical AI safety and governance/strategy tracks, aiming to build the pipeline of professionals who can work at AI safety organizations and reduce existential risk from advanced AI.
| Metric | Value |
|---|---|
| Score | 1.54e-09 |
| Potential Positive Impact | 0.1000 |
| Feasibility | 0.0126 |
| Novelty | 0.3000 |