Anthropic新研究：Claude自主对齐自身，超越28位研究员方案
New Anthropic research shows Claude aligning Claude all by itself.
做对齐和递归自我改进研究的同学必看，Claude 自主发现的安全训练方法已超越资深研究员方案，建议深读原文了解其闭环流程与门控设计。
New Anthropic research shows Claude aligning Claude all by itself.
Anthropic的新研究显示，Claude完全自主地使Claude自身对齐。
And here the model being corrected was more capable than the model correcting it
而在这里，被纠正的模型比进行纠正的模型能力更强。
Looks like another great research at the frontier of recursive self-improvement. AI is reaching the point where it can do much of the research required to make its own stronger successors safer.
这看起来又是递归自我改进前沿的另一项伟大研究。AI正达到一个阶段，即它能够承担大部分研究任务，这些任务旨在使其自身更强大的后继者更加安全。
The loop it describes closes on itself: propose, train, score, repeat, no researcher required.
它描述的循环自我闭合：提出、训练、评分、重复，无需研究人员介入。
Claude autonomously discovered safety-training methods that outperformed one-shot ideas from 28 experienced researchers.
Claude自主发现的安全训练方法，其表现超越了28位经验丰富的研究人员一次性提出的想法。
Five agents worked in parallel for up to 48 hours, sharing results while hidden tests and capability gates screened out overfitting and obvious regressions.
五个代理并行工作长达48小时，共享结果，同时隐藏测试和能力门槛筛选出过拟合和明显的性能退化。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力