Alignment Forum and LessWrong
Posts from the Alignment Forum, and LessWrong posts on AI safety.
60 of 5,395 documents, newest first
Where is all the AI safety research uplift?Near HorizonsAjeya: AI is dramatically accelerating AI research right nowElegy for What’s Left of the NormalNotes on Take-Off Speeds and RSIFive Types of “Corrigibility Projects” I Don’t Want to FundWe urgently need to practice before we can pauseMETR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026People Should Resolve Tensions In AI Safety ResearchOpenAI: Astra covertly underperforms on safety researchPositive and Negative AI FuturesDesiderata for the Theory of AgentsAgainst strawmanning unilateral pausesOrganizations I'd recommend for fundingAutomating subtle emergent misalignment researchTeaching Qwen WhySome “coherence arguments” are two different things in a trenchcoatAdditional Course Context (for AI Safety Fundamentals; unofficial)Scott Alexander’s Against SURA Seems Deeply ConfusedTowards The AutograntorExplaining “AI might obliterate you” to senior policymakers in IndiaAgainst SURAMicrodosing the spirit of AnthropicLeaking Muse's Charter, Meta's model specYou can now "@Grok" on LessWrongNotes on “We’re in Triage Mode for AI Safety” by Miles BrundageScattered thoughts on how doing math feels these daysSome Memories of Early AnthropicA cat-belling problem for FDTAnthropic are the Good GuysFrom overeliciting to undereliciting LLM propensities in just a few monthsSetting the Scene (Public v1; October 2026)San Francisco State University in TwilightPatching reward hacks won’t fix reward hackingWah wah wah wah do more AI R&DGreater good arguments in AI safetyMonitorability evaluations and policy proposalsNeural nets are exceptionally vulnerable to goal self-hackWhy I’m becoming a professor in AI safety in BeijingA brief(ish) introduction to LM psychologyThere Will Be HellsACS Infra: free access to large base models for LLM psychology researchhuggingface.wtf, an accessible but complete explanationCapital Ownership Will Make You Go Crazy FirstThe Alignment Problem: Computer Science Meets Analytic PhilosophyI Did This Without Your KnowledgeNeural networks are not already cleanDeconfusing CorrigibilityAstra and Sol can take low probability actions with no CoTPreface to the Computational Agent Theory and ASI Alignment SequenceA Simpler and Better Debate Protocol for Scalable OversightLinkpost: Neural Networks Are Already CleanYour prompt… plus everything you've ever (publicly) writtenThe Continual Constitutional Convention: How language models revise the documents that govern themPersonas in a Pre-Internet Language ModelResearch Lessons I Learned at MATS & RedwoodOpenAI Can't Self-Police Because It Fires the People Raising the AlarmLabs’ current alignment plans seem terribleMore people should write basic AI safety explainersHello World, AI Doompop