Founders

— Cofounding teams —

Francisco Carvalho (xiq)

A Global Serendipity Layer

Full ideaTwitter works as a global serendipity layer routing opportunity and sensemaking. I've gotten most of my funding, friends, jobs, and relationships through twitter. However, twitter wasn't explicitly designed for this. The GSL will increase serendipity 100x by letting you ambiently coordinate with weak ties. Weak ties are the main source of serendipity. You have vastly more diverse friends of friends than in your direct circle. (1973, Granovetter, The Strength of Weak Ties https://www.jstor.org/stable/2776392) We can leverage the potential of weak ties. By connecting multiple data sources from a user (twitter, chats, notes), an agent can infer intents and share them with a user's extended network (e.g. within cuties.app) to find relevant opportunities and information. There are four questions between us and the GLS: 1. Ingestion: how to ingest abundant and timely data about a user? 2. Privacy and taste: how to reveal semi-private information in a way that you would endorse? can the agent represent your interests and bring you back the best stuff? 3. Network architecture and bootstrapping: what's the protocol the agents use to talk? (protocols like my project https://claudeconnect.io) how do users network? (high trust directories of people like https://cuties.app) 4. Opportunity mining: how to efficiently search over the network and match intents? A few components that we've already worked on (research): 1. a model of user taste and values that can inform much better content recommendations; (obtained by hill climbing this eval xiqo.substack.com/p/agentic-taste-modeling-lab-notes); 2. a digital "switchboard operator" that understands the graph of people's models of each other's interest and competence per topic; (obtained by hill climbing this eval https://docs.google.com/document/d/1BW6Wi0qrIV8bGmXgjbViyQtyK6p-FWBQK815jeOkMWQ); 3. opportunity mining: given a model of taste and of the social graph, we can discover intents latent in people's public writing (twitter, substack) and eventually private writing (messages, journals) and use match them to actualize more potential. (studied here https://xiqo.substack.com/p/opportunity-mining-lab-notes-6)Twitter works as a global serendipity layer routing opportunity and sensemaking. I've gotten most of my funding, friends, jobs, and relationships through twitter. However, twitter wasn't explicitly designed for this. The GSL will increase serendipity 100x by letting you ambiently coordinate with weak ties. Weak ties are the main source of serendipity. You have vastly more diverse friends of friends than in your direct circle. (1973, Granovetter, The Strength of Weak Ties https://www.jstor.org/stable/2776392) We can leverage the potential of weak ties. By connecting multiple data sources from a user (twitter, chats, notes), an agent can infer intents and share them with a user's extended network (e.g. within cuties.app) to find relevant opportunities and information. There are four questions between us and the GLS: 1. Ingestion: how to ingest abundant and timely data about a user? 2. Privacy and taste: how to reveal semi-private information in a way that you would endorse? can the agent represent your interests and bring you back the best stuff? 3. Network architecture and bootstrapping: what's the protocol the agents use to talk? (protocols like my project https://claudeconnect.io) how do users network? (high trust directories of people like https://cuties.app) 4. Opportunity mining: how to efficiently search over the network and match intents? A few components that we've already worked on (research): 1. a model of user taste and values that can inform much better content recommendations; (obtained by hill climbing this eval xiqo.substack.com/p/agentic-taste-modeling-lab-notes); 2. a digital "switchboard operator" that understands the graph of people's models of each other's interest and competence per topic; (obtained by hill climbing this eval https://docs.google.com/document/d/1BW6Wi0qrIV8bGmXgjbViyQtyK6p-FWBQK815jeOkMWQ); 3. opportunity mining: given a model of taste and of the social graph, we can discover intents latent in people's public writing (twitter, substack) and eventually private writing (messages, journals) and use match them to actualize more potential. (studied here https://xiqo.substack.com/p/opportunity-mining-lab-notes-6)

Other ideas1) Discourse graphs for science on ATProto; 2) Personal daemons (in the vein of Gwern's Guardian Angels) connecting p2p (like my project https:/claudeconnect.io), given a directory of people (like https:/cuties.app). 3) A public-facing website doing a digital anthropology of postrationalist twitter and its cultural influence. A data-driven + first-person account. Following work here https://xiqo.substack.com/p/discovering-the-postrat-canon-in 4) A "switchboard operator" twitter-bot that knows everyone and tags them under relevant tweets, be they questions, opportunities, jokes. 5) A catalogue of niche high taste twitter bots QTing high-taste content in specific domains like AI alignment, longevity, semiconductors, geopolitics - potentially for arbitrary topics. Following work on https://bangers.community-archive.org/, which found methods of aggregating the highest signal tweets from a community.

Christine Shiba

☞ x.com/christineist

Cultivate a hyperstition engine on Cuties!

✦ cuties.app

Full ideaContext: Cuties! is a community-based social app for open minded and intellectually curious people. We already have an existing brand with high trust, quality people, and good vibes. Communities include lesswrong, the embassy, feytopia, vibecamp, fractal, edge city, and more. In ~2 years it has led to 2K sign ups, ~900 MAU, engagements and a baby. Given our high trust environment, I think there is opportunity to conduct social ai experimentation. Our users will be more forgiving and more willing to engage in our explorations. Idea: Cuties AI will be a community agent that explores stated and ambient intentions and finds opportunities latent within the user base. Many users have already written in detail about how they want to live their lives on their Cuties profiles. Cuties AI can capitalize on this existing repository to get started; however we will also develop an in-product loop to encourage users to 1) state their intentions, and 2) to refine and restate their intentions based on the quality of opportunities found. By doing this we will teach both the agent and our users to increase their capacity to surface serendipity, and to improve their abilities to bring dreams into existence through well stated intentions.

Other ideas- introduce "spirit daemons" into Cuties profiles as a way to set up your personal agent. This will allow for experimenting with feed, channels, etc with agents as equal and soulful characters in our society. - a vouch network that acts as community infrastructure. This explores the idea of Cuties as an open social app, where you own your identification card, your vouches, your membership in certain communities, etc. - tpot wiki where we memorialize the transformative and memetic ideas that created this subculture https://docs.google.com/document/d/1zHITIv4L_a75XeE0Yq09BmBEoYOBZtd00ir1tkIYG5E/edit?tab=t.0#heading=h.ajgdt9ncvzx7

Joey Bream

☞ linkedin.com/in/joseph-bre…

Customer verification for the biotech industry

Dr. Phil Palmer

☞ safely.bio

Customer verification for the biotech industry

✦ safely.bio

Full ideaAI is making it easier to engineer biology. As increasingly bio-capable models become open source, the number of bad actors capable of causing serious harm grows rapidly. Model safeguards will not hold. The remaining defensible layer is the physical bio supply chain: an attacker still needs synthetic DNA to turn a design into a real threat. Securing DNA synthesis is the most promising chokepoint (screendna.org). At safely.bio, we are building Verify: customer verification software that DNA synthesis providers integrate into their intake to run know-your-customer (KYC) checks on the biologists placing orders. Our recent benchmark shows Verify outperforms the current state-of-the-art AI customer screening tool (preprint coming shortly). We have had 30+ customer conversations and are sprinting to close design-partner pilots ahead of incoming regulation (US S.3741, EU Biotech Act). Longer term, this becomes a security layer for the biotech industry, comparable to what cybersecurity is today.

Other ideasWe’re open to exploring other ideas at the intersection of AI and biosecurity. Once AI agents are everywhere, we need to defend against new risks: - Benchtop DNA synthesiser security: Portable synthesisers, similar to 3D printers, remove the chokepoint by allowing anyone to print dangerous sequences. On-device controls to prevent this are extremely neglected. - Cloud lab security. Automated AI-native wet-labs are scaling up and set to absorb lots of requests from agents. We would create the verification layer that detects dangerous processes

Cecilia Roos*

☞ apps.apple.com/us/story/id…

Easy online truth-seeking for epistemic resilience

✦ cherry-tree.netlify.app

Full ideaTruth is a precondition for making informed choices about the most important things in our lives: health, families, livelihoods, our democracy. When people can’t tell what’s true, they fall back on fear, gut feeling, or the most compelling story. As it stands today, ordinary people lack practical tools to defend themselves. We want to change this. We’re building tools to empower people to seek the truth and converge on a shared set of facts. Our first tool gives people an in-context, on-demand way to check claims anywhere they go on the web. With one click, our tool separates verifiable claims from opinion or prediction; gives concise, plain-language evidence; and quotes/links to high-quality primary sources.

Other ideasBuild more tools for epistemic resilience in the age of rapidly advancing AI.

Charlie Giattino*

☞ linkedin.com/in/charliegia…

Empower people to find truth and think better

✦ cherry-tree.netlify.app

Full ideaMost people care about the truth. But in a world increasingly flooded with information, it’s becoming harder to find. So people are turning to LLMs: one study found that of the millions of requests X users sent Grok and Perplexity, nearly 8% asked whether something was true (Renault, Mosleh & Rand, 2026). But people need better help — and we plan to build it for them. Grok and other LLMs often sound authoritative but hide their reasoning, mix fact with opinion, and hallucinate information and sources. We're building an inline, on-demand truth-finding tool that does the opposite: it reasons transparently, separates fact from opinion, and quotes & links directly to trustworthy sources so users can easily verify and dig deeper. Along the way, it helps people get better at finding the truth themselves. This is just the first step. We want to build a suite of tools that help increase society’s epistemic resilience in the age of AI.

Other ideasBuild more tools for epistemic resilience in the age of rapidly advancing AI.

Sophia Wang*

☞ sophiajwang.com

Live maps and forecasts of AI power concentration

✦ x.com/mapping_ai/status/20…

Full ideaWe already built and run Mapping AI (mapping-ai.org), an open, interactive map of the US AI policy landscape, with [1,600+] entities and [3,000+] relationships covering who is shaping AI governance, what they believe about regulation and timelines, and how they're funded and connected. At Surplus we would extend Mapping AI into an interactive, live forecasting tool of various AI power concentration scenarios and pair it with a public blog, working name [Civic Compact]. The extension adds a scenario layer to the existing map: working with Metaculus's networked-forecasting platform Radiant, we build on AI 2027's scenario-based approach and expand it to several power-concentration world models (gradual disempowerment, authoritarian consolidation, capital concentration, among others), each rendered as a directed causal map with continuously updating probabilities from domain experts and professional forecasters, grounded in the entity and relationship data the tool already maintains. [Civic Compact] carries the insights to a general audience through narrative pieces that follow the scenarios as the maps and forecasts update, and it feeds readers back into the tool as users and contributors. Both would build on infrastructure we already run (Mapping AI database, agent and verification harnesses, and the interactive visualizations), and leverage a broad network of orgs + people who have been excited about supporting Mapping AI already.

Other ideasThe interactive research blog will range wider than the power-concentration scenarios, covering ideas that come out of this work and our other projects in the space. Two pieces are in progress already: one is on economic measurement systems for tracking AI-driven labor displacement, drawing on components like register-based statistics in the Nordics and Estonia, India's biometric consent-based data aggregation, and privacy-preserving computation between labs and statistical agencies; we have proposed a pilot computing an employment indicator jointly between a frontier lab and a statistical agency without either side exposing raw data. The other is a review of corporate governance mechanisms and proposals, spanning new legal forms, financial instruments, disclosure, and more, to increase the capacity for shareholder activism in the age of frontier AI labs; it reviews the existing history of corporations and the public benefit corporation structure, then proposes mechanisms with real public accountability. Planned after those: (1) innovative schemes to get AI-literate people into government, covering incentive systems, the legal pathways that exist to raise compensation, and what USDS and DOGE set as precedent; (2) an interview series with candidates who lost their primaries, on their own diagnosis of what went wrong, focused on AI and tech policy platforms and their experience with lobbying organizations; (3) investigating why for-profit LLCs make it easier to fundraise than non-profits in the AI safety space and how to leverage that (a meta-analysis of the surplus.dev structure), (4) etc.

Anushree Chaudhuri*

☞ anushreec.substack.com/about

Live maps and forecasts of AI power concentration

✦ mapping-ai.org

Full ideaWe already built and run Mapping AI (mapping-ai.org), an open, interactive map of the US AI policy landscape, with 1,600+ entities and 3,000+ relationships covering who is shaping AI governance, what they believe about regulation and timelines, and how they're funded and connected. At Surplus we would extend Mapping AI into an interactive, live forecasting tool of various AI power concentration scenarios and pair it with a public blog, working name Civic Compact (https://civiccompact.org). The extension adds a scenario layer to the existing map: working with Metaculus's networked-forecasting platform Radiant, we build on AI 2027's scenario-based approach and expand it to several power-concentration world models (gradual disempowerment, authoritarian consolidation, capital concentration, among others), each rendered as a directed causal map with continuously updating probabilities from domain experts and professional forecasters, grounded in the entity and relationship data the tool already maintains. Civic Compact carries the insights to a general audience through narrative pieces that follow the scenarios as the maps and forecasts update, and it feeds readers back into the tool as users and contributors. Both would build on infrastructure we already run (Mapping AI database, agent and verification harnesses, and the interactive visualizations), and leverage a broad network of orgs + people who have been excited about supporting Mapping AI already. See Civic Compact for working drafts of two initial publications: https://civiccompact.org/

Other ideasThe interactive research blog will range wider than the power-concentration scenarios, covering ideas that come out of this work and our other projects in the space. Two pieces are in progress already: one is on economic measurement systems for tracking AI-driven labor displacement, drawing on components like register-based statistics in the Nordics and Estonia, India's biometric consent-based data aggregation, and privacy-preserving computation between labs and statistical agencies; we have proposed a pilot computing an employment indicator jointly between a frontier lab and a statistical agency without either side exposing raw data (“Measuring the AI economy” https://civiccompact.org/measuring-the-ai-economy). The other is a review of corporate governance mechanisms and proposals, spanning new legal forms, financial instruments, disclosure, and more, to increase the capacity for shareholder activism in the age of frontier AI labs; it reviews the existing history of corporations and the public benefit corporation structure, then proposes mechanisms with real public accountability (“Whose company is it?” https://civiccompact.org/corporations). Planned after those: (1) innovative schemes to get AI-literate people into government, covering incentive systems, the legal pathways that exist to raise compensation, and what USDS and DOGE set as precedent; (2) an interview series with candidates who lost their primaries, on their own diagnosis of what went wrong, focused on AI and tech policy platforms and their experience with lobbying organizations/PACs; (3) investigating why for-profit LLCs make it easier to fundraise than non-profits in the AI safety space and how to leverage that (a meta-analysis of the surplus.dev structure), (4) etc.

— Solo founders —

Owen Shen

☞ linkedin.com/in/owenshen

AI that proactively knows ur preferences

Full ideaOkay so tbqh I don't really know yet. I previously thought about ways to extend curation (in a world filled w/ more slop and more content, you need to rely more on filtering the info) in a way that is trained on your own personal preferences. Then a friend linked Gwern's Guardian Angels post and it was like a more developed version of what I had been thinking. Anyway, here's the original idea best as I can describe it, without being too influenced: train a model locally/isolated off of all your daily browsing history / keyboard history to better understand you. The goal is somewhat a mix between a replacement for every For You page in that it surfaces content recommendations, but is, importantly, **adjustable**. If you want to see more recipes, or your goal for the week is to drill more Spanish, it can better surface content for that. At the same time, you're providing feedback on how good the content being surfaced actually is so the model can predict how you might respond to pieces of content, improving the filtering flow. The end goal is an automation that proactively reads content, decides how you'd react to it, and then surfaces ones that it predicts you'd end up wanting to view.

Other ideasAm open to other types of software that helps with codifying human preferences into digital form.

Theo Ryzhenkov

☞ home.theor.net

Curius (https://curius.app/), executed perfectly

Full ideaCurius is an app that provides community notes -like experience across the web. You can save web pages and comment on them, and your network, if visiting them, can see your comments. I want to improve on it drastically, providing an end-it-all tool. For example, by making a local application, I will expose a localhost daemon to subscribe to my OpenAPI API, for arbitrary scripting: for example, by piping newly saved pages with a news tag through an LLM agent, it can verify sources, find alternative coverage, etc.; or I can write a batteries-included module for integration with Personal Knowledge Management systems, like Notion, Obsidian or Logseq. I will implement closed groups, so that you can save and comment not only publicly, but for selected people. For example, if you are founding something, you can comment inline on your competitors pages for your co-founders. I will make a discovery system that, with machine learning, analyses what signals make you rate pages high, and recommends new articles, posts, pages, across your friend network. I will make paid annotations, and attract famous people to annotate some websites for me to advertise this feature. For example, someone could annotate this form with their advice on a background check of Surplus evaluators and what might impress them, and put it behind paywall. We will take a small cut off of all transactions. Etc. I will make everything extremely fast, seamless, keyboard driven and multiplatform. I have an experience with all of this for Posthaste (https://github.com/theoryzhenkov/posthaste).

Other ideasI might want to make an online, selective program for young talented folk specifically to help them orient themselves in life, and help them out with relocation, money, education. This program would have access to prestigious endorsements for its students, would prioritise helping students out, and not just pushing them higher on the the "success ladder", and discovery of unrealised talent, like potentially "rescuing" people from bad environments, providing legal help for talented teenagers in bad households, etc. I could collaborate with FABRIC (https://www.fabric.camp/), with whom I am working already.

Aniket Panjwani

☞ youtube.com/@aniketapanjwani

Unlock Restricted Data for AI Assisted Research

Full ideaI propose creating a content-led AI implementation and training organization for social science/policy organizations. The central piece of content will be a clear guide on how to use local LLMs for research purposes with restricted data, e.g. with census data or tax data. The guide will enable researchers working with restricted data to know for their use cases which combination of local LLM and GPU infrastructure they ought to use. The guide will be a living document, and it will also contain evals for social science/policy relevant applications. The guide will exist in several formats. First, dedicated articles on some website. Second, a calculator and/or chat UI which people can use to map their local compute needs to a procurement decision and personalized local LLM roadmap (similar to PC Part Picker - https://pcpartpicker.com/). Third, a series of YouTube videos educating users on local LLMs and demonstrating physically how to put together the local LLM rig. As a result of reading this guide, the motivated user and/or their IT team should be able to make a GPU procurement decision and set up local LLMs for their use cases. In practice, research centers will still look for help with implementation/nuanced procurement decisions and staff training. So, the revenue model will be to do consulting with the high quality free content as advertising attracting ideal customers (policy organizations, central banks, government institutions, law firms) - similar to the business model of Every (https:/every.to).

Other ideasAnother idea I may want to pursue is AI-based software for causal inference. Researchers in both academia and the private sector need to credibly establish a wide variety of causal effects. In the private sector, one can often rely on causal identification via randomization from tightly scoped experiments. But in many other contexts one may only have quasi-experimental or observational data with which to establish causation. For example, an organization like Coefficient Giving often needs to make decisions off messy, real-world data using quasi-experimental techniques like difference-in-differences. However, techniques like difference in differences themselves branch off in nuanced ways as aspects of a quasi-experimental setting satisfy certain assumptions and violate others. When researchers attempt to establish causal estimates in these quasi-experimental settings, they have to carefully thread their situation to the appropriate technique or set of techniques to use. I want to create software which automates this mapping of "quasi-experimental setting/context" to "exact technique or set of techniques" to use. Such a software will both allow researchers to better use causal techniques, and allow non-researchers to use causal techniques (similar to how Elicit allows non-researchers to produce high-level pharmaceutical research). As a PoC, I'm working on developing an agent which formalizes this practitioner's guide to difference-in-differences (https://www.aeaweb.org/articles?from=f&id=10.1257%2Fjel.20251650). I think a revenue model with an open source core + SaaS upsell could work. This idea could also be something which bundles broadly with the main "Unlocking Restricted Data for AI Assisted Research" idea I'm proposing above. To see this, consider how Every's popular "Compound Engineering" plugin (https://github.com/EveryInc/compound-engineering-plugin) acts as advertising for its own consulting services. Similarly, a "causal inference" Claude/Codex plugin could act as advertising for my organization's consulting offers.Another idea I may want to pursue is AI-based software for causal inference. Researchers in both academia and the private sector need to credibly establish a wide variety of causal effects. In the private sector, one can often rely on causal identification via randomization from tightly scoped experiments. But in many other contexts one may only have quasi-experimental or observational data with which to establish causation. For example, an organization like Coefficient Giving often needs to make decisions off messy, real-world data using quasi-experimental techniques like difference-in-differences. However, techniques like difference in differences themselves branch off in nuanced ways as aspects of a quasi-experimental setting satisfy certain assumptions and violate others. When researchers attempt to establish causal estimates in these quasi-experimental settings, they have to carefully thread their situation to the appropriate technique or set of techniques to use. I want to create software which automates this mapping of "quasi-experimental setting/context" to "exact technique or set of techniques" to use. Such a software will both allow researchers to better use causal techniques, and allow non-researchers to use causal techniques (similar to how Elicit allows non-researchers to produce high-level pharmaceutical research). As a PoC, I'm working on developing an agent which formalizes this practitioner's guide to difference-in-differences (https://www.aeaweb.org/articles?from=f&id=10.1257%2Fjel.20251650). I think a revenue model with an open source core + SaaS upsell could work. This idea could also be something which bundles broadly with the main "Unlocking Restricted Data for AI Assisted Research" idea I'm proposing above. To see this, consider how Every's popular "Compound Engineering" plugin (https://github.com/EveryInc/compound-engineering-plugin) acts as advertising for its own consulting services. Similarly, a "causal inference" Claude/Codex plugin could act as advertising for my organization's consulting offers.

Haoxing Du

☞ haoxingdu.com

Consumer Reports for AI model behavior

✦ docs.google.com/document/d…

Full ideaAI labs and third-party evaluators have been focused on evaluations of *capabilities*. In contrast, the *behavior* of AI models is not well characterized, or even known, to the developers—traits such as: How often does a model agree with a user in cases where the user is factually wrong (sycophancy)? Does it deny having done something it actually did when asked directly (gaslighting)? Does it behave differently when it believes it is unmonitored versus monitored (covert behavior)? AI models of the same capability level can have quite different behaviors or personalities, which affects users' experience interacting with them, and in turn how they will shape the world. And some of these behaviors are relevant to safety and catastrophic risk, e.g. evaluation awareness, sandbagging, and lying. If models are misaligned, this type of behavioral evaluation could and would generate direct evidence for it. I want to build the thing (website? organization? startup?) that measures this: Consumer Reports / Wirecutter for AI models, or a field guide / ethogram for models. METR but for behaviors.

Other ideasMaybe something along the lines of an AI-powered service for cheap, rapid scientific replications

Derik Kauffman

☞ derikk.com

Roast my post and fact checking for newsrooms

Full ideaAI is clearly good enough to fact check verifiable claims. Yet major newspapers still make mistakes that prove they’re not using it in the proofreading and fact checking process. Just two months ago, a NYT headline called NATO the “North American Treaty Organization”. Building on the open source but now abandoned Roast my Post, I want to make a tool for editors and reporters to check pieces before publication, improving journalistic quality and bolstering the information commons.

Other ideasI filled out a second application with my other idea. Also, I'm pretty interested in bringing to life a number of technologies in https://www.forethought.org/research/design-sketches-for-a-more-sensible-world, would love to discuss with people working on that stuff

Vaishnav Sunil

☞ optimaloutliers.com/p/liqu…

Predictive hiring signal in the age of noise

✦ cloutcareers.com ✦ optimaloutliers.com/p/star…

Full ideaLLMs significantly worsened the problem of noise in scalable hiring channels. As they help generate plausible applications at scale, they void any signal in artefacts like cover letters. As channels get noisier, adverse selection becomes more of a concern, feeding back on itself. I was confident early that this would only get worse. Clout works on both the sourcing and evaluation sides of recruiting. On sourcing, we started exclusively via qualified referrals: instead of posting a job on the internet, we go to our proprietary network and ask people we trust "who do you think we should hire," and we compensate second-degree referrals to make it more generative. (I've written more about the mechanism here: https://www.optimaloutliers.com/p/startup-hiring-solving-the-problem) On evaluation, I've been working on more consulting-like contracts with clients — mostly in AI safety — doing everything across the stack, from building work tests to conducting reference checks, using LLMs tactically but pairing them with a shared high-context human. The piece I'm currently testing in a more focussed way with one of my clients is AI resistant takehome (as traditional take-homes have started getting one shotted) which is potentially scalable.

Other ideasThere are several ideas within the talent /recruiting space that are worth exploring, I will list these in the order of familiarity to me: Trust-weighted referral network AI resistant takehome Company Chatbots that can quickly assess the cognitive capacity and fit of a new applicant by asking open ended questions. Systematic reference checks (human + voice bot) Infrastructure that helps source talent by connecting to islands of signal before they disappear (substacks, podcasts, meetup groups etc)

Beat Hagenlocher

☞ beathagenlocher.com

An additional front door to EA

Full ideaI think EA techniques could have a way, way larger audience, but currently don't have that because of the way they're presented. In my mind, the process into EA should work like the following: 1. Someone talks about something that annoys them in the world or they worry about (say: climate change) 2. An EA friend of theirs jumps in and is like: In fact, you _can_ do something about that. 3. Then, they hand over a website that has their specific problem preselected (let's say climate change), and then the website (as an interactive text) walks them through the options they have, depending on the resources (skills, network, money) they bring with them, and risk appetite they have. Along the way (of starting to give money at all; of choosing a better charity to give money to; of brainstorming a career change, ...) it teaches EA mental models _when they become relevant_, and creates a learning journey (with inline spaced repetition cards for understanding) out of them. 4. They learn and progress, hopefully identify with EA themselves, and give these resources to other people. Success! But that's not the case right now. Currently, you can only point them to: - EA meetups (which almost never address them where they are, which is at Climate Change or wherever) - the EA website (which gives lots of high-level justifications for EA and EA content—but doesn't address _their problem_ either) - or the 80k website (where you _may_ be getting somewhere with your questions, but the answers are less "this is what you can do", but rather: "Sorry, that's not important (enough), we can't help you"). And I think that all of these three alternatives make us lose many, many people. The website I'm proposing should fill that gap: 1. It should address EA newbies where they are, make them feel understood, help them make their specific next actions more effective, and then make them take more and more EA-aligned actions, until they know all of the basics. 2. They will only be able to take more and more of those EA-aligned actions if they semi-regularly visit the website, and that (I think) will only happen when there's enough repeatedly useful content on there. For that, I propose workbook guides that people can work through when: - they need to do a career change - they need to decide where their priorities lie, and what they want to do with their life - they need to find a new job - they want to evaluate whether they are 'still on track' with a plan they've made To put this together: This should both work well as a grayspace for beginners (which EA also doesn't have right now) and as a workbook for intermediate/advanced EAs to continually come back to (whenever there is something like, say, a career change, again)I think EA techniques could have a way, way larger audience, but currently don't have that because of the way they're presented. In my mind, the process into EA should work like the following: 1. Someone talks about something that annoys them in the world or they worry about (say: climate change) 2. An EA friend of theirs jumps in and is like: In fact, you _can_ do something about that. 3. Then, they hand over a website that has their specific problem preselected (let's say climate change), and then the website (as an interactive text) walks them through the options they have, depending on the resources (skills, network, money) they bring with them, and risk appetite they have. Along the way (of starting to give money at all; of choosing a better charity to give money to; of brainstorming a career change, ...) it teaches EA mental models _when they become relevant_, and creates a learning journey (with inline spaced repetition cards for understanding) out of them. 4. They learn and progress, hopefully identify with EA themselves, and give these resources to other people. Success! But that's not the case right now. Currently, you can only point them to: - EA meetups (which almost never address them where they are, which is at Climate Change or wherever) - the EA website (which gives lots of high-level justifications for EA and EA content—but doesn't address _their problem_ either) - or the 80k website (where you _may_ be getting somewhere with your questions, but the answers are less "this is what you can do", but rather: "Sorry, that's not important (enough), we can't help you"). And I think that all of these three alternatives make us lose many, many people. The website I'm proposing should fill that gap: 1. It should address EA newbies where they are, make them feel understood, help them make their specific next actions more effective, and then make them take more and more EA-aligned actions, until they know all of the basics. 2. They will only be able to take more and more of those EA-aligned actions if they semi-regularly visit the website, and that (I think) will only happen when there's enough repeatedly useful content on there. For that, I propose workbook guides that people can work through when: - they need to do a career change - they need to decide where their priorities lie, and what they want to do with their life - they need to find a new job - they want to evaluate whether they are 'still on track' with a plan they've made To put this together: This should both work well as a grayspace for beginners (which EA also doesn't have right now) and as a workbook for intermediate/advanced EAs to continually come back to (whenever there is something like, say, a career change, again)

Other ideasI'm excited about at least half of the ideas you lined out, and would be happy with being a technical/design cofounder of another startup in Surplus: I'm excited about AI guardian angels, better knowledge bases, AI for better decision making, basically any other impactful and fun public facing website projects (if I think I can contribute to them well). I'd be especially excited about working on any of the infrastructure projects: Specifically the open source S-Process and an EA talent platform sound intriguing—but since I haven't worked a lot with funding and grants yet, I might be a bit less of a fit.

Nathan Young*

☞ x.com/NathanpmYoung

10x Community Notes across different media

✦ goodheart-labs.github.io/c…

Full ideaCommunity Notes are the bit of context that shows up next to a claim, instead of the claim being deleted or changed. They work sustainably on X. They are more trusted than factchecks [1], can scale quickly, and our AI notewriter writes them at the same helpfulness level as the typical human. It is expected that social media companies do fact-checking, but what about other areas? Community Notes works really well, hence it's being taken up by Meta and TikTok, but we don't have the same expectations for podcasts, news, journals, government outputs and more. I am building pilots to demonstrate the demand for this product, with a view to changing the norms. Community Notes provides content that most users think is valuable and helps deal with misleading content without algorithmic suppression or censorship. I want to see that in more places. [1] https://academic.oup.com/pnasnexus/article/3/7/pgae217/7686087

Other ideasForecasting tools, mostly. I build forecasting bots (Metaculus, Manifold) and a Polymarket trading bot. I think there ought to be a way to provide high quality scenario analysis to companies. I don't think anyone has cracked how to deliver cheap, accurate forecasts in a form that company leadership are interested in.