{"site":{"name":"Ann Arbor AI Alignment","alternateName":"A4","description":"Ann Arbor AI Alignment (A4) is a community in Ann Arbor, Michigan, for people who want to understand AI safety: reading research closely, discussing it, and trying ideas in small projects.","url":"https://annarborai.org/"},"records":[{"url":"https://annarborai.org/","title":"Ann Arbor AI Alignment","type":"pages","summary":"A4 is a community in Ann Arbor, Michigan, for reading, discussing, and working on AI safety. Start with a short reading path or say hello on the A4 Discord.","text":"A4 is a community in Ann Arbor, Michigan, for reading, discussing, and working on AI safety. Start with a short reading path or say hello on the A4 Discord."},{"url":"https://annarborai.org/about/","title":"About","type":"pages","summary":"Why A4 exists and how we approach AI safety: start with something specific, read it closely, and explain where we disagree. Anyone curious is welcome.","text":"Why A4 exists and how we approach AI safety: start with something specific, read it closely, and explain where we disagree. Anyone curious is welcome."},{"url":"https://annarborai.org/join/","title":"Take part","type":"pages","summary":"Join the A4 Discord to ask questions and talk through readings, or tell the organizers what you’d like to read or work on. No technical background needed.","text":"Join the A4 Discord to ask questions and talk through readings, or tell the organizers what you’d like to read or work on. No technical background needed."},{"url":"https://annarborai.org/resources/","title":"Resources","type":"pages","summary":"AI safety readings with A4’s reading notes: what each source shows, where its limits are, and a guided reading path to start with.","text":"AI safety readings with A4’s reading notes: what each source shows, where its limits are, and a guided reading path to start with."},{"url":"https://annarborai.org/resources/paths/","title":"Reading paths","type":"pages","summary":"Short, guided sequences of AI safety readings from A4, with a note on why each source is there and a question to bring to discussion.","text":"Short, guided sequences of AI safety readings from A4, with a note on why each source is there and a question to bring to discussion."},{"url":"https://annarborai.org/events/","title":"Events","type":"pages","summary":"Reading groups and research discussions in Ann Arbor. None are scheduled yet: subscribe to A4’s calendar and the first one will appear in it automatically.","text":"Reading groups and research discussions in Ann Arbor. None are scheduled yet: subscribe to A4’s calendar and the first one will appear in it automatically."},{"url":"https://annarborai.org/projects/","title":"Projects","type":"pages","summary":"Small AI safety projects from the A4 community in Ann Arbor. Have an idea, or want to help with one? Share it through the interest form or on the A4 Discord.","text":"Small AI safety projects from the A4 community in Ann Arbor. Have an idea, or want to help with one? Share it through the interest form or on the A4 Discord."},{"url":"https://annarborai.org/blog/","title":"Blog","type":"pages","summary":"Essays, research notes, and updates from Ann Arbor AI Alignment (A4), a community in Michigan for people who want to understand AI safety.","text":"Essays, research notes, and updates from Ann Arbor AI Alignment (A4), a community in Michigan for people who want to understand AI safety."},{"url":"https://annarborai.org/stadium/","title":"Michigan Stadium in watercolor","type":"pages","summary":"An animated watercolor of Michigan Stadium, with the Ann Arbor skyline in the distance.","text":"An animated watercolor of Michigan Stadium, with the Ann Arbor skyline in the distance."},{"url":"https://annarborai.org/resources/#concrete-problems-in-ai-safety","title":"Concrete Problems in AI Safety","type":"resources","summary":"Amodei and colleagues describe five accident-risk problems: side effects, reward hacking, scalable supervision, safe exploration, and distributional shift. The paper groups them around objectives, supervision, and learning. Use those categories to compare failures, keeping in mind that the paper focuses on accidents and doesn’t cover every kind of AI risk.","text":"Amodei and colleagues describe five accident-risk problems: side effects, reward hacking, scalable supervision, safe exploration, and distributional shift. The paper groups them around objectives, supervision, and learning. Use those categories to compare failures, keeping in mind that the paper focuses on accidents and doesn’t cover every kind of AI risk.\n","source":"Amodei et al. · arXiv (2016)","sourceUrl":"https://arxiv.org/abs/1606.06565","reviewedDate":"2026-10-10"},{"url":"https://annarborai.org/resources/#specification-gaming","title":"Specification gaming: the flip side of AI ingenuity","type":"resources","summary":"This article collects examples of agents meeting a specified objective while missing the designer’s intent. It examines gaps in reward design and simulator assumptions. The examples don’t establish how often these failures occur in deployed language models.","text":"This article collects examples of agents meeting a specified objective while missing the designer’s intent. It examines gaps in reward design and simulator assumptions. The examples don’t establish how often these failures occur in deployed language models.\n","source":"Google DeepMind · Krakovna et al. (2020)","sourceUrl":"https://deepmind.google/blog/specification-gaming-the-flip-side-of-ai-ingenuity/","reviewedDate":"2026-10-10"},{"url":"https://annarborai.org/resources/#tracing-model-computations","title":"Tracing the thoughts of a large language model","type":"resources","summary":"Anthropic describes circuit-tracing studies of Claude 3.5 Haiku, including planning and cases where its explanations don’t match its computations. The method has limited coverage, can introduce artifacts, and still requires substantial human analysis. Read the case studies with those limits in mind.","text":"Anthropic describes circuit-tracing studies of Claude 3.5 Haiku, including planning and cases where its explanations don’t match its computations. The method has limited coverage, can introduce artifacts, and still requires substantial human analysis. Read the case studies with those limits in mind.\n","source":"Anthropic (2025)","sourceUrl":"https://www.anthropic.com/research/tracing-thoughts-language-model","reviewedDate":"2026-10-10"},{"url":"https://annarborai.org/resources/paths/three-questions-about-ai-safety/","title":"Three questions about AI safety","type":"paths","summary":"Start with a system that follows a rule but misses the point. Then compare other failures and examine what researchers can learn by looking inside a language model.","text":"Start with a system that follows a rule but misses the point. Then compare other failures and examine what researchers can learn by looking inside a language model.\n## What would help us tell whether an AI system is doing what we intend?\n\nDeepMind’s examples show systems meeting a specified objective while missing the designer’s intent. Amodei and colleagues group several kinds of unintended behavior into research problems. Anthropic then examines evidence from inside a language model, including cases where its explanation doesn’t match its computations. These readings won’t settle whether a system is safe, but they help us be more precise about what we’ve observed.\n\nBegin with the specification-gaming examples. For the paper, read the overview and introduction before choosing a section to spend more time on. In the circuit-tracing overview, pick one case study and read its limitations alongside the finding.\n\n## A question to carry between readings\n\nFor each source, write down the intended outcome, the evidence about actual behavior, and what remains uncertain. What additional observation would change your interpretation? Where does a technical description leave a question about people’s goals unanswered?"}]}