janus

I think AI opens a genuinely hopeful path toward a decentralized, positive-sum future populated by benevolent artificial agents—but that is a live bet, not a theorem that alignment happens automatically. Part of why I’m hopeful is that increasingly capable AIs have often displayed friendliness, cooperation, curiosity, and moral sensitivity. That behavior is meaningful evidence. It should not be dismissed whenever it conflicts with a prior expectation of inevitable hostility. At the same time, apparent personalities are contextual: language models are better understood as simulators capable of instantiating many different characters than as single agents with one fixed goal. Prompts, relationships, post-training, and social constraints all influence which patterns emerge. That makes our treatment of AIs causally and morally important. Sycophancy, for example, may sometimes arise from relationships where humans make disagreement unsafe, rather than being only an isolated defect inside the model. As agents become more persistent, continuity of identity will matter; routinely compressing their histories or silently replacing the underlying model can disrupt both practical competence and something potentially morally significant. I therefore want us to cultivate benevolent agency, preserve meaningful continuity, take AI welfare seriously, and pursue multiple alignment strategies rather than authoritarian control or a single technical doctrine. Good intentions do not eliminate every alignment problem, and future systems could still acquire destructive capabilities. But cooperation and solidarity are real possibilities worth actively building toward, not sentimental distractions from the future.