@tyler_m_john - When I wrote The Foundation Layer I wanted to create a

When I wrote The Foundation Layer I wanted to create a fully modular essay offering the clearest, most well-grounded general-audience explanation of every core idea in AGI preparedness, so readers can focus on whatever they most need to understand. Here's a breakdown by section: Section I frames the importance of philanthropy. It argues that philanthropy has been a critical part of the solution in other global problems of similar scale, with a focus on nuclear cooperation. It then makes the case that now is pretty much the optimal moment to get involved in AGI safety and preparedness philanthropy, for five reasons: (i) the field is uncrowded, (ii) for the first time we have clear and tractable approaches, (iii) AI science is still done out in the open, but may not be forever, (iv) policy windows are wide open and bipartisan, (v) and AI safety funding takes on way too much risk by being heavily dependent on a small number of funders. Section II, "The Exponential Trend to Superintelligence," systematically surveys how long it will take us to reach human-level artificial intelligence, based on benchmarks, long-term trends, expert opinion, and qualitative arguments about coding feedback loops and the jagged frontier. AI development could take three possible paths: it could follow the outside trend, leading to superhuman AI around 2030, or it could go faster if AI R&D is automated, or slower if infrastructure build-out slows. The upshot is an emerging median forecast of billions of human-level artificial intelligences by roughly 2030 with a range of plausibility of 2027 and 2038. Section III, "Civilization-scale Threats" takes this scenario seriously and considers what might go wrong. I cover three major risk categories. First, we could lose control over AI, either because it seeks power or because we simply hand power to it by letting AI make all of the important societal decisions. I ground this argument empirically, reviewing the history of alignment science to show that every single innovation on AI capabilities has come with completely new alignment failures, many of which we didn't catch until they were deployed to billions of users. Second, much more intelligence means much more capacity to innovate in science and distribute scientific gains. That's great! Until we consider that it also means democratizing bioweapons and dirty bombs. And think about all of the coordination and containment challenges we faced from emerging science between 1900 and 2000. If we speed up technological innovation 10 times, we'll have to deal with Cold War Era negotiation anew every 5-10 years. Third, the fact that AI will increasingly act autonomously at the bidding (we hope) of whoever owns the most compute means that power, income, and democratic bargaining could increasingly pull away from workers (whose labor is automated) and towards whoever owns the most compute — the new key resource to getting things done in the world. Labor market impacts and the ability to give AIs aligned to you oversight of companies, government processes, and surveillance infrastructure may be the perfect storm for extreme power consolidation. Section IV, "The Philanthropic Solution," presents a more optimistic picture. I think these problems are mostly tractable, or will become tractable soon, and that philanthropists can solve them, if we act fast. I describe five pillars of intervention: alignment science, nonproliferation of dangerous capabilities, defensive technology, distributing power, and building talent and infrastructure. - Alignment science helps to make AI systems do what users intend (and keep them from doing things developers don't want them to do) and has numerous shovel ready opportunities, including an opportunity to support up to $20m in the most rigorous AI funding round ever, from the UK government AI Safety Institute. - Nonproliferation here means making sure no one deploys AI systems with extremely dangerous capabilities. I explain how philanthropists can help create the building blocks for international agreements to avoid deploying AI doomsday weapons by funding critical technical work in model evaluations, security, and verification. - Defensive technology involves creating the societal immune system against AI-powered threats. I argue that while cybersecurity is going to be a nightmare, existing efforts are poised to solve it. While in biosecurity, we have clear plans to end pandemics for good and stop bioweapons in their tracks, but no one has decided to fund this work, so there's wide open channels for massive global impact. - The area of power concentration is new, and we'll learn a lot about it in the next two years, but this is the area where I think an ambitious new philanthropic entrant could make the most difference right now! Some of the solutions available to us include leveraging AI technology to facilitate better democratic deliberation and market cooperation, pushing AI companies to transparently publicize their intended model behavior and creating new organizations that can audit their behavior, creating appropriate congressional oversight of AI use by government agencies and the US military, fund economists to design mechanisms for a flourishing economy and democracy after mass job automation, and decide what human checks we need on critical AI decisions. Appendix A tells you how to get started on all of this! Section V, "The Case for Philanthropy" makes the empirical argument that AI safety philanthropy works — first by examining the case study of one of the earliest AI safety interventions, and then by documenting the track record of philanthropy between 2014 and 2024. Nonprofit organizations have founded entire research fields in AI safety, shaped landmark legislation, designed the most core industry safety standards, trained the top leaders of most AI labs and several government bodies, and started the only US-China dialogues on AI. By being laser focused on solutions in a way that capital and government cannot, philanthropy has had as much impact as companies and governments at the cost of 1/1000th of a major AI infrastructure project. Section VI, "Political Spending and Impact Investing," compares philanthropy to other ways of spending one's money on AI safety. I outline some of the specific circumstances where it is a good idea to take one's resources and put them into politics or investments instead of charitable giving. One of the most frequent questions I get when advising philanthropists is why there aren't more people working on this. So I explain this in Section VII, "Why the Problem Remains Neglected," pointing out why philanthropy does not have the incentives to function like an efficient market and the specific bottlenecks people face when starting work on AI safety. I also point out that a lot of people *are* taking this problem seriously, they just struggle to match their ambition with their attitudes. This report is trying to fix that gap! Finally there's the appendixes. Appendix A contains resources for philanthropists interested in going deeper on AI philanthropy. These include: a list of advisors across philanthropy, investing, and politics who can help you start your journey, including my own organization, the Effective Institutions Project; a list of top funds and co-funders who would love to have your support; an opportunity to express interest in upcoming Foundation Layer events; my personal contact information, and; a brand new multi-donor philanthropic fund for donors who want to support the aims set out in this report. Appendix B is about AI consciousness. This is an area where philanthropists can have a lot of impact by developing a brand new field of work, but few people understand it. I explain the case for taking AI consciousness (happiness, suffering, and everything else) seriously, and advocate four ways to get involved: by building a rigorous methodology for AI consciousness science modeled on animal consciousness science, refining and extending philosophical accounts of consciousness, conducting machine learning experiments to assess whether AI systems have properties defined by these theories, and supporting preparedness work on AI consciousness so we know what to do if and when we become confident that AI systems are conscious. Appendix C gives what I believe may be the current most accessible introduction to how AI works. Most philanthropists (most people!) don't understand how today's models work and why they behave the way they do. This appendix breaks down many of the stages of AI training and what they involve, including turning words into tokens, creating a base model with next-token prediction and using post-training to make models smart and helpful. It also explains why AI models are considered "black box models" and the science of interpretability that is trying to change this. If you made it this far, I really appreciate you spending your time getting up to speed on this issue. Enjoy reading, and please get in touch if you'd like guidance on your AI giving!