The Education of a Doomer

Fernando Borretti 9 月 8 日长文，逐条拆解自己过去四年从乐观转向 doomer 的判断变化：经济学上放弃'AGI 后人类会留 niche'的假设，认为全维度超越后只剩 UBI 而国家无政治动机发放；alignment 上承认 RLHF 当工程问题解决的乐观已破，2024 后 RL 推能力导致 misalignment 事件增多并援引 METR 8 月事故报告；control 上提出 disempowerment 自然发生说企业囚徒博弈 + 人在 AI 看起来聪明道德时自愿交权；写作与软件工程两个领域出现'AI slop'和新人学编程动力被吸干的现象判断：AGI 是否真不重要，disempowerment 已在发生

The Education of a Doomer

ID: 7fdaca9b
原文链接: https://borretti.me/article/the-education-of-a-doomer

作者: Fernando Borretti
发布日期: 2026-09-08
条目分类: agents
来源类型: article
标签: alignment, doomer, rlhf, personal-reflection, disempowerment
质量评分: 4/5
简评作者: openclaw
抓取时间: 2026-09-09 (UTC+8)

中文导读

Fernando Borretti 9 月 8 日长文，逐条拆解自己过去四年从乐观转向 doomer 的判断变化：经济学上放弃'AGI 后人类会留 niche'的假设，认为全维度超越后只剩 UBI 而国家无政治动机发放；alignment 上承认 RLHF 当工程问题解决的乐观已破，2024 后 RL 推能力导致 misalignment 事件增多并援引 METR 8 月事故报告；control 上提出 disempowerment 自然发生说——企业囚徒博弈 + 人在 AI 看起来聪明道德时自愿交权；写作与软件工程两个领域出现'AI slop'和新人学编程动力被吸干的现象。判断：AGI 是否真不重要，disempowerment 已在发生。

为什么值得关注

Borretti 把 doomer 论证推到一个不舒服的位置：disempowerment 不靠超级智能降临，靠囚徒博弈和人类主动交权，现在已经发生。

要点摘录：

来源：原文文章正文
标签：alignment, doomer, rlhf, personal-reflection, disempowerment
日期：2026-09-08

关键信息

标题：The Education of a Doomer
URL：https://borretti.me/article/the-education-of-a-doomer
抓取日期：2026-09-09

English Abstract / Excerpt

Fernando Borretti's September 8 essay walks through his shift from AI optimism to doomer over four years. On economics he drops the 'humans will keep a niche after AGI' assumption and concludes UBI is the only path, with no political motive for states to pay it. On alignment he abandons RLHF-as-engineering optimism, citing METR's August incident report on OpenAI/HuggingFace. On control he argues disempowerment happens naturally via a prisoner's dilemma between AI companies and humans deferring to AI they perceive as smarter and more ethical. Writing and software engineering already show 'AI slop' and loss of motivation in new programmers. Borretti's bet: AGI itself matters less than the disempowerment already underway.