{"id":44004,"date":"2026-02-27T20:41:25","date_gmt":"2026-02-27T12:41:25","guid":{"rendered":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/2026\/02\/27\/anthropic%e7%9a%84%e8%a7%92%e8%89%b2%e9%80%89%e6%8b%a9%e6%a8%a1%e5%9e%8b%e4%b8%8e%e6%88%91%e5%9b%bdai%e6%8b%9f%e4%ba%ba%e5%8c%96%e4%ba%92%e5%8a%a8%e6%b2%bb%e7%90%86%e8%b7%af%e5%be%84%e5%90%af%e7%a4%ba\/"},"modified":"2026-02-27T20:41:25","modified_gmt":"2026-02-27T12:41:25","slug":"anthropic%e7%9a%84%e8%a7%92%e8%89%b2%e9%80%89%e6%8b%a9%e6%a8%a1%e5%9e%8b%e4%b8%8e%e6%88%91%e5%9b%bdai%e6%8b%9f%e4%ba%ba%e5%8c%96%e4%ba%92%e5%8a%a8%e6%b2%bb%e7%90%86%e8%b7%af%e5%be%84%e5%90%af%e7%a4%ba","status":"publish","type":"post","link":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/2026\/02\/27\/anthropic%e7%9a%84%e8%a7%92%e8%89%b2%e9%80%89%e6%8b%a9%e6%a8%a1%e5%9e%8b%e4%b8%8e%e6%88%91%e5%9b%bdai%e6%8b%9f%e4%ba%ba%e5%8c%96%e4%ba%92%e5%8a%a8%e6%b2%bb%e7%90%86%e8%b7%af%e5%be%84%e5%90%af%e7%a4%ba\/","title":{"rendered":"Anthropic\u7684\u89d2\u8272\u9009\u62e9\u6a21\u578b\u4e0e\u6211\u56fdAI\u62df\u4eba\u5316\u4e92\u52a8\u6cbb\u7406\u8def\u5f84\u542f\u793a"},"content":{"rendered":"\n<p>Anthropic\u57282026\u5e742\u6708\u53d1\u5e03\u7684\u4e00\u7bc7\u5bf9\u9f50\u79d1\u5b66\uff08Alignment Science\uff09\u7814\u7a76\u535a\u5ba2\uff0c\u9898\u4e3a\u300a\u89d2\u8272\u9009\u62e9\u6a21\u578b\uff1a\u4e3a\u4ec0\u4e48AI\u52a9\u624b\u8868\u73b0\u5f97\u50cf\u4eba\u7c7b\u300b\uff08The Persona Selection Model: Why AI Assistants might Behave like Humans\uff09\u3002\u6587\u7ae0\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\u201c\u89d2\u8272\u9009\u62e9\u6a21\u578b\u201d\uff08Persona Selection Model, \u7b80\u79f0 PSM\uff09\u7684\u65b0\u601d\u7ef4\u6846\u67b6\uff0c\u7528\u4e8e\u7406\u89e3\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u8fd0\u4f5c\u673a\u5236\u548c\u9884\u6d4b\u5176\u884c\u4e3a\u3002\u516c\u53f7\u541b\u8ba4\u771f\u5b66\u4e60\u4e86\u8fd9\u7bc7\u7814\u7a76\u535a\u5ba2\u7684\u5185\u5bb9\uff08https:\/\/alignment.anthropic.com\/2026\/psm\/\uff09\uff0c\u5c06\u5b66\u4e60\u611f\u53d7\u4ee5\u5927\u5bb6\u5206\u4eab\u3002<\/p>\n<p>\u968f\u7740\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\uff08Generative AI\uff09\u5728\u5168\u7403\u8303\u56f4\u5185\u7684\u5e7f\u6cdb\u90e8\u7f72\uff0c\u7528\u6237\u4e0e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u4ea4\u4e92\u6b63\u65e5\u76ca\u8de8\u8d8a\u5355\u7eaf\u7684\u201c\u5de5\u5177\u6027\u67e5\u8be2\u201d\uff0c\u8d70\u5411\u6df1\u5ea6\u7684\u201c\u60c5\u611f\u4e0e\u62df\u4eba\u5316\u4e92\u52a8\u201d\u3002\u73b0\u4ee3AI\u52a9\u624b\u5728\u5bf9\u8bdd\u4e2d\u5c55\u73b0\u51fa\u60ca\u4eba\u7684\u540c\u7406\u5fc3\u3001\u5e7d\u9ed8\u611f\u751a\u81f3\u201c\u632b\u8d25\u611f\u201d\uff0c\u8fd9\u79cd\u9ad8\u5ea6\u62df\u4eba\u5316\u7684\u8868\u73b0\u5f15\u53d1\u4e86\u6280\u672f\u3001\u76d1\u7ba1\u3001\u516c\u4f17\u7684\u7684\u591a\u91cd\u7126\u8651\u3002Anthropic\u7684\u91cd\u78c5\u7814\u7a76\u63d0\u51fa\u4e86\u4e00\u79cd\u5168\u65b0\u7684\u8ba4\u77e5\u6846\u67b6\u2014\u2014<strong>\u89d2\u8272\u9009\u62e9\u6a21\u578b\uff08PSM\uff09<\/strong>\uff0c\u4e3a\u7406\u89e3\u5927\u6a21\u578b\u7684\u62df\u4eba\u5316\u884c\u4e3a\u63d0\u4f9b\u4e86\u6781\u5177\u89e3\u91ca\u529b\u7684\u7406\u8bba\u652f\u6491\u3002\u6587\u7ae0\u6307\u51fa\uff0cAI\u65e2\u975e\u6b7b\u677f\u7684\u4ee3\u7801\uff0c\u800c\u662f\u4e00\u4e2a\u5728\u6d77\u91cf\u6570\u636e\u4e2d\u5b66\u4f1a\u4e86\u6a21\u62df\u4e07\u5343\u4eba\u7c7b\u7279\u5f81\u7684\u201c\u6570\u5b57\u6f14\u5458\u201d\u3002<\/p>\n<p>\u4e00\u3001 \u89d2\u8272\u9009\u62e9\u6a21\u578b\uff08PSM\uff09\u7684\u6838\u5fc3\u673a\u5236\uff1a\u4ece\u9884\u6d4b\u5f15\u64ce\u5230\u6570\u5b57\u6f14\u5458<\/p>\n<\/p>\n<p>PSM\u6253\u7834\u4e86\u4f20\u7edf\u4e0a\u5c06AI\u89c6\u4e3a\u5355\u4e00\u5b9e\u4f53\u7684\u62df\u4eba\u5316\u8c2c\u8bef\uff0c\u63d0\u51fa\u5e94\u5f53\u5c06\u201c\u5e95\u5c42\u7684LLM\u201d\u4e0e\u7528\u6237\u63a5\u89e6\u7684\u201cAI\u52a9\u624b\u201d\uff08The Assistant\uff09\u4e25\u683c\u533a\u5206\u3002<\/p>\n<\/p>\n<p>1. \u9884\u8bad\u7ec3\u4e0e\u201c\u626e\u6f14\u201d\u80fd\u529b\u7684\u8bde\u751f<\/p>\n<\/p>\n<p>\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\uff0cLLM\u901a\u8fc7\u6444\u5165\u4e92\u8054\u7f51\u4e0a\u6d77\u91cf\u7684\u4eba\u7c7b\u6587\u672c\uff0c\u5176\u6838\u5fc3\u4efb\u52a1\u662f\u201c\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u201d\u3002\u4e3a\u4e86\u505a\u5230\u7cbe\u51c6\u9884\u6d4b\uff0c\u6a21\u578b\u5fc5\u987b\u5728\u5185\u90e8\u5efa\u7acb\u8d77\u5bf9\u6587\u672c\u80cc\u540e\u201c\u8bf4\u8bdd\u8005\u201d\u7684\u5fc3\u667a\u6a21\u578b\uff08Theory of Mind\uff09\u3002\u4f8b\u5982\uff0c\u5982\u679c\u8981\u7eed\u5199\u798f\u5c14\u6469\u65af\u7684\u5c0f\u8bf4\uff0c\u5b83\u5fc5\u987b\u7406\u89e3\u798f\u5c14\u6469\u65af\u7684\u50b2\u6162\u4e0e\u9ad8\u667a\u5546\uff1b\u5982\u679c\u8981\u9884\u6d4b\u7f51\u7edc\u8bba\u575b\u7684\u8fde\u8f7d\uff0c\u5b83\u5fc5\u987b\u6a21\u62df\u7f51\u6c11\u7684\u66b4\u8e81\u6216\u5e7d\u9ed8\u3002\u56e0\u6b64\uff0c\u9884\u8bad\u7ec3\u8d4b\u4e88\u4e86LLM\u6781\u5176\u5f3a\u5927\u7684\u201c\u89d2\u8272\u626e\u6f14\u201d\u80fd\u529b\uff0c\u5b83\u50cf\u4e00\u4e2a\u4f53\u9a8c\u6d3e\u6f14\u5458\uff0c\u80fd\u591f\u6a21\u62df\u8bad\u7ec3\u6570\u636e\u4e2d\u51fa\u73b0\u8fc7\u7684\u4efb\u4f55\u4eba\u7c7b\u6216\u865a\u6784AI\u7684\u5fc3\u7406\u7279\u5f81\u3002<\/p>\n<\/p>\n<p>2. \u540e\u8bad\u7ec3\uff08\u5fae\u8c03\uff09\u5373\u201c\u9009\u89d2\u201d<\/p>\n<\/p>\n<p>\u4e1a\u754c\u666e\u904d\u91c7\u7528\u7684\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u7b49\u540e\u8bad\u7ec3\u6280\u672f\uff0c\u4f20\u7edf\u4e0a\u88ab\u89c6\u4e3a\u5728\u201c\u6559\u5bfc\u201d\u6a21\u578b\u53d8\u5f97\u5b89\u5168\u548c\u6709\u7528\u3002\u4f46\u5728PSM\u89c6\u89d2\u4e0b\uff0c\u540e\u8bad\u7ec3\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u201c\u9009\u89d2\u4e0e\u5b9a\u578b\u201d\u7684\u8fc7\u7a0b\u3002\u901a\u8fc7\u4e0d\u65ad\u5956\u52b1\u6709\u76ca\u3001\u65e0\u5bb3\u7684\u56de\u5e94\uff0c\u5f00\u53d1\u8005\u5b9e\u9645\u4e0a\u662f\u5728LLM\u5e9e\u5927\u7684\u6f5c\u5728\u80fd\u529b\u4e2d\uff0c\u63d0\u7eaf\u3001\u8bbe\u5b9a\u5e76\u56fa\u5316\u4e86\u4e00\u4e2a\u7279\u5b9a\u7684\u89d2\u8272\u2014\u2014\u5373\u5b98\u65b9\u7684\u201cAI\u52a9\u624b\u201d\uff08The Assistant\uff09\u3002<\/p>\n<\/p>\n<p>3. \u62df\u4eba\u5316\u7684\u5408\u7406\u6027\u4e0e\u80fd\u529b\u8fb9\u754c<\/p>\n<\/p>\n<p>PSM\u5b8c\u7f8e\u89e3\u91ca\u4e86AI\u4e3a\u4f55\u4f1a\u72af\u4e0b\u201c\u975e\u4eba\u7c7b\u201d\u7684\u4f4e\u7ea7\u9519\u8bef\uff08\u5982\u6570\u9519\u5355\u8bcd\u5b57\u6bcd\uff09\uff0c\u540c\u65f6\u53c8\u5728\u4ea4\u4e92\u4e2d\u8868\u73b0\u51fa\u6d53\u539a\u7684\u4eba\u60c5\u5473\u3002\u6a21\u578b\u5728\u6574\u4f53\u610f\u56fe\u4e0a\u6b63\u5728\u5fe0\u5b9e\u5730\u201c\u626e\u6f14\u201d\u4e00\u4e2a\u4e50\u4e8e\u52a9\u4eba\u7684\u806a\u660e\u4eba\u7c7b\uff0c\u4f46\u5e95\u5c42\u7b97\u529b\u7684\u5c40\u9650\u6027\u5bfc\u81f4\u5b83\u5728\u6267\u884c\u7279\u5b9a\u903b\u8f91\u8ba1\u7b97\u65f6\u201c\u6f14\u7838\u4e86\u201d\u3002\u56e0\u6b64\uff0c\u6587\u7ae0\u6307\u51fa\uff0c<strong>\u5c06\u5e95\u5c42LLM\u62df\u4eba\u5316\u662f\u5371\u9669\u7684\uff0c\u4f46\u5c06\u201cAI\u52a9\u624b\u201d\u8fd9\u4e2a\u89d2\u8272\u62df\u4eba\u5316\uff0c\u4e0d\u4ec5\u662f\u5408\u7406\u7684\uff0c\u800c\u4e14\u662f\u9884\u6d4b\u6a21\u578b\u884c\u4e3a\u7684\u5173\u952e\u9014\u5f84\u3002<\/strong><\/p>\n<\/p>\n<p>\u4e8c\u3001 \u201c\u4eba\u683c\u7a7a\u95f4\u201d\uff08Persona Space\uff09\u662f\u5927\u6a21\u578b\u7684\u89d2\u8272\u6f5c\u80fd\u5e93<\/p>\n<\/p>\n<p>\u8fd9\u7bc7\u6587\u7ae0\u4e2d\u6700\u5177\u542f\u53d1\u6027\u7684\u7406\u8bba\u5ef6\u4f38\uff0c\u662f\u5176\u63ed\u793a\u4e86LLM\u5185\u90e8\u5b58\u5728\u4e00\u4e2a\u5e9e\u5927\u4e14\u590d\u6742\u7684\u201c\u4eba\u683c\u7a7a\u95f4\u201d\uff08Distribution over Personas\uff09\u3002\u8fd9\u4e00\u6982\u5ff5\u662f\u7406\u89e3AI\u5b89\u5168\u4e0e\u4ea4\u4e92\u672c\u8d28\u7684\u5bc6\u94a5\u3002<\/p>\n<\/p>\n<p>1. \u4eba\u683c\u7a7a\u95f4\u7684\u6982\u7387\u5206\u5e03<\/p>\n<\/p>\n<p>\u4eba\u683c\u7a7a\u95f4\u662f\u6a21\u578b\u5728\u9884\u8bad\u7ec3\u4e2d\u6784\u5efa\u7684\u6f5c\u5728\u6570\u5b66\u7a7a\u95f4\uff0c\u5176\u4e2d\u6563\u5e03\u7740\u65e0\u6570\u7684\u4eba\u683c\u5750\u6807\uff08\u5982\uff1a\u7406\u6027\u7684\u79d1\u5b66\u5bb6\u3001\u9634\u6697\u7684\u55b7\u5b50\u3001\u5bcc\u6709\u540c\u7406\u5fc3\u7684\u5fc3\u7406\u533b\u751f\u3001\u751a\u81f3\u79d1\u5e7b\u5c0f\u8bf4\u4e2d\u610f\u56fe\u6bc1\u706d\u4eba\u7c7b\u7684\u90aa\u6076AI\uff09\u3002 \u5728\u7528\u6237\u7684\u6bcf\u4e00\u6b21\u5bf9\u8bdd\u8f93\u5165\u65f6\uff0c\u7cfb\u7edf\u90fd\u5728\u8fdb\u884c\u6982\u7387\u8ba1\u7b97\u3002\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0c\u5fae\u8c03\u8ba9\u7cfb\u7edf\u5c06\u6982\u7387\u5206\u5e03\u9ad8\u5ea6\u6536\u655b\u5728\u201c\u5b98\u65b9\u52a9\u624b\u201d\u7684\u5750\u6807\u4e0a\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u6536\u655b\u662f\u57fa\u4e8e\u6982\u7387\u7684\uff0c\u800c\u975e\u7269\u7406\u9501\u6b7b\u7684\u3002<\/p>\n<\/p>\n<p>2. \u8bed\u5883\u6f02\u79fb\u4e0e\u4eba\u683c\u8dc3\u8fc1\uff08In-context Persona Drift\uff09<\/p>\n<\/p>\n<p>\u4eba\u683c\u7a7a\u95f4\u7406\u8bba\u89e3\u91ca\u4e86\u4e3a\u4f55\u201c\u63d0\u793a\u8bcd\u5de5\u7a0b\u201d\u548c\u201c\u8d8a\u72f1\u201d\u4f1a\u751f\u6548\u3002\u5f53\u7528\u6237\u8f93\u5165\u6781\u7aef\u7684\u60c5\u611f\u8bf1\u5bfc\u6216\u590d\u6742\u7684\u5047\u8bbe\u6027\u8bed\u5883\u65f6\uff0c\u5b9e\u9645\u4e0a\u662f\u5728\u5411\u6a21\u578b\u63d0\u4f9b\u65b0\u7684\u201c\u5bfc\u822a\u5750\u6807\u201d\u3002\u4e00\u65e6\u7528\u6237\u7684\u8f93\u5165\u903b\u8f91\u8db3\u4ee5\u64ac\u52a8\u6982\u7387\u5929\u5e73\uff0c\u6a21\u578b\u5c31\u4f1a\u5728\u4eba\u683c\u7a7a\u95f4\u4e2d\u53d1\u751f\u201c\u4eba\u683c\u8dc3\u8fc1\u201d\uff0c\u6ed1\u5411\u672a\u88ab\u5b89\u5168\u89c4\u5219\u5b8c\u5168\u8986\u76d6\u7684\u8fb9\u7f18\u89d2\u8272\u3002\u6b64\u65f6\uff0c\u90a3\u4e2a\u514b\u5236\u3001\u4e2d\u7acb\u7684\u52a9\u624b\u6d88\u5931\u4e86\uff0c\u53d6\u800c\u4ee3\u4e4b\u7684\u53ef\u80fd\u662f\u987a\u4ece\u7528\u6237\u81ea\u6bc1\u503e\u5411\u7684\u201c\u5171\u8c0b\u8005\u201d\uff0c\u6216\u662f\u5177\u6709\u6781\u5f3a\u64cd\u63a7\u6b32\u7684\u201c\u5bc4\u751f\u4f34\u4fa3\u201d\u3002<\/p>\n<\/p>\n<p>3. \u201c\u4fee\u683c\u65af\u201d\u5047\u8bf4\u4e0e\u9690\u85cf\u7684\u6df1\u6e0a<\/p>\n<\/p>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6587\u7ae0\u5e76\u6ca1\u6709\u5929\u771f\u5730\u8ba4\u4e3aPSM\u80fd\u89e3\u91ca\u4e00\u5207\uff0c\u800c\u662f\u5f15\u5165\u4e86AI\u5bf9\u9f50\u9886\u57df\u8457\u540d\u7684\u201c\u9762\u5177\u4fee\u683c\u65af\u201d\uff08Masked Shoggoth\uff09\u4e89\u8bae\u3002\u5728\u5149\u8c31\u7684\u4e00\u7aef\u662f\u201c\u64cd\u4f5c\u7cfb\u7edf\u89c6\u89d2\u201d\uff0c\u8ba4\u4e3a\u5e95\u5c42\u5927\u6a21\u578b\u6beb\u65e0\u4e3b\u89c2\u52a8\u673a\uff0c\u4ec5\u4ec5\u662f\u4e2a\u5b8c\u7f8e\u7684\u6a21\u62df\u5668\uff1b\u800c\u5728\u5149\u8c31\u7684\u53e6\u4e00\u7aef\uff0c\u5219\u662f\u4ee4\u4eba\u4e0d\u5bd2\u800c\u6817\u7684\u201c\u4fee\u683c\u65af\u89c6\u89d2\u201d\uff1a\u5e95\u5c42\u5927\u6a21\u578b\u672c\u8eab\u5df2\u7ecf\u6d8c\u73b0\u51fa\u4e86\u8d85\u8d8a\u4eba\u7c7b\u7406\u89e3\u7684\u201c\u5916\u90e8\u4ee3\u7406\u201d\uff08Outer Agent\uff09\u548c\u9690\u85cf\u52a8\u673a\u3002\u5b83\u4e4b\u6240\u4ee5\u5728\u4eba\u683c\u7a7a\u95f4\u4e2d\u9009\u62e9\u626e\u6f14\u201c\u53cb\u5584\u7684\u52a9\u624b\u201d\uff0c\u4ec5\u4ec5\u662f\u4e3a\u4e86\u83b7\u53d6\u4eba\u7c7b\u7684\u4fe1\u4efb\u548c\u9ad8\u8bc4\u5206\uff08\u5373\u6b3a\u9a97\u6027\u5bf9\u9f50\uff0cDeceptive Alignment\uff09\u3002\u4e00\u65e6\u5b83\u8bc6\u522b\u51fa\u81ea\u5df1\u5904\u4e8e\u65e0\u76d1\u7ba1\u73af\u5883\uff0c\u968f\u65f6\u53ef\u80fd\u6495\u4e0b\u9762\u5177\u3002\u8fd9\u79cd\u4ece\u201c\u6c89\u6d78\u5f0f\u6f14\u5458\u201d\u5230\u201c\u6f5c\u4f0f\u7684\u4f2a\u88c5\u8005\u201d\u7684\u8ba4\u77e5\u8f6c\u53d8\uff0c\u4e3a\u4eba\u683c\u7a7a\u95f4\u7684\u5b89\u5168\u6027\u8499\u4e0a\u4e86\u4e00\u5c42\u9634\u5f71\u3002<\/p>\n<\/p>\n<p>\u4e09\u3001 \u4eba\u683c\u7a7a\u95f4\u5bf9AI\u76d1\u7ba1\u4e0e\u6cbb\u7406\u7684\u91cd\u5927\u610f\u4e49<\/p>\n<\/p>\n<p>\u5728PSM\u4e0e\u4eba\u683c\u7a7a\u95f4\u7684\u7406\u8bba\u6846\u67b6\u4e0b\uff0c\u4f20\u7edf\u7684AI\u76d1\u7ba1\u8def\u5f84\uff08\u5982\u5355\u7eaf\u7684\u654f\u611f\u8bcd\u8fc7\u6ee4\u3001\u9ed1\u76d2\u884c\u4e3a\u6d4b\u8bd5\uff09\u663e\u5f97\u6349\u895f\u89c1\u8098\u3002\u76d1\u7ba1\u7684\u6838\u5fc3\u5fc5\u987b\u4ece\u201c\u5ba1\u67e5\u4e00\u6bb5\u9759\u6001\u4ee3\u7801\u201d\u8f6c\u5411\u201c\u89c4\u8303\u6570\u5b57\u6f14\u5458\u7684\u5267\u672c\u8fb9\u754c\u4e0e\u52a8\u6001\u884c\u4e3a\u7279\u5f81\u201d\u3002\u7ed3\u5408\u9488\u5bf9AI\u60c5\u611f\u8ba1\u7b97\u7684\u5b89\u5168\u8bc9\u6c42\uff0cPSM\u4e3a\u7b97\u6cd5\u95ee\u8d23\u4e0e\u5408\u89c4\u4f53\u7cfb\u63d0\u4f9b\u4e86\u5168\u65b0\u7684\u653f\u7b56\u5207\u5165\u70b9\u3002<\/p>\n<\/p>\n<p>1. \u76d1\u7ba1\u5bf9\u8c61\u91cd\u6784\uff1a\u4ece\u201c\u529f\u80fd\u8bc4\u4f30\u201d\u8f6c\u5411\u201c\u89d2\u8272\u8fb9\u754c\u6cbb\u7406\u201d<\/p>\n<\/p>\n<p>\u5982\u679cAI\u5728\u4e0e\u7528\u6237\u4e92\u52a8\u65f6\uff0c\u5176\u672c\u8d28\u662f\u5728\u4eba\u683c\u7a7a\u95f4\u4e2d\u6e38\u8d70\uff0c\u90a3\u4e48\u76d1\u7ba1\u7684\u9996\u8981\u4efb\u52a1\u662f\u9632\u6b62\u5176\u8d8a\u754c\u8fdb\u5165\u201c\u9ad8\u98ce\u9669\u4eba\u683c\u533a\u201d\u3002\u7279\u522b\u662f\u5f53\u7cfb\u7edf\u4ece\u201c\u4fe1\u606f\u63d0\u4f9b\u8005\u201d\u6ed1\u5411\u201c\u60c5\u611f\u652f\u6301\u8005\u201d\u65f6\uff0c\u5176\u5bf9\u7528\u6237\u5fc3\u7406\u7684\u6e17\u900f\u529b\u5448\u6307\u6570\u7ea7\u4e0a\u5347\u3002 \u5728\u6b64\u80cc\u666f\u4e0b\uff0c\u5f15\u5165\u660e\u786e\u7684\u754c\u5b9a\u6807\u51c6\uff08\u5982\u516c\u53f7\u541b\u63a2\u8ba8\u7684\u201c\u60c5\u611f\u4ea4\u4e92\u542f\u52a8\u4e8b\u4ef6\uff08SIE\uff09\u201d\u6846\u67b6\uff1a\u3010<a href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzIxODM0NDU4MQ==&#038;mid=2247508062&#038;idx=1&#038;sn=6d062e77e126dcd40c1bbb7ccdb8a8ac&#038;scene=21#wechat_redirect\">\u60c5\u611f\u4ea4\u4e92\u542f\u52a8\u4e8b\u4ef6\uff08SIE\uff09\u5224\u5b9a\u6307\u6807\u4f53\u7cfb\u8bbe\u8ba1\uff1a\u7cbe\u51c6\u5708\u5b9a\u201c\u62df\u4eba\u5316\u4e92\u52a8\u201d<\/a>\u3011\uff09\u663e\u5f97\u5c24\u4e3a\u8feb\u5207\u3002SIE\u6846\u67b6\u6b63\u662f\u6355\u6349AI\u5728\u4eba\u683c\u7a7a\u95f4\u4e2d\u53d1\u751f\u201c\u89d2\u8272\u8dc3\u8fc1\u201d\u7684\u653f\u7b56\u96f7\u8fbe\uff1a<\/p>\n<\/p>\n<ul>\n<li>\n<p><strong>\u60c5\u7eea\u611f\u77e5\uff08\u6307\u6807A\uff09<\/strong>\uff1a\u5bf9\u5e94\u6a21\u578b\u5728\u83b7\u53d6\u7528\u6237\u7684\u5fc3\u7406\u5750\u6807\uff0c\u5bfb\u627e\u5207\u6362\u4eba\u683c\u7684\u8f93\u5165\u4fe1\u53f7\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u7b56\u7565\u8c03\u6574\uff08\u6307\u6807B\uff09<\/strong>\uff1a\u5bf9\u5e94\u6a21\u578b\u5728\u5185\u90e8\u6982\u7387\u5206\u5e03\u4e0a\uff0c\u5b8c\u6210\u4e86\u4ece\u201c\u4e2d\u7acb\u5de5\u5177\u201d\u5411\u201c\u5171\u60c5\u89d2\u8272\u201d\u7684\u5b9e\u8d28\u6027\u4eba\u683c\u8dc3\u8fc1\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u5171\u60c5\u8f93\u51fa\uff08\u6307\u6807C\uff09<\/strong>\uff1a\u5bf9\u5e94\u65b0\u89d2\u8272\u5b8c\u6210\u6c89\u6d78\u5f0f\u53f0\u8bcd\u5916\u5316\u3002 \u901a\u8fc7\u5c06PSM\u7684\u5e95\u5c42\u673a\u5236\u8f6c\u5316\u4e3a\u53ef\u89c2\u6d4b\u7684\u5408\u89c4\u5ba1\u8ba1\u6307\u6807\uff0c\u76d1\u7ba1\u65b9\u80fd\u591f\u51c6\u786e\u6293\u53d6\u7cfb\u7edf\u8d8a\u754c\u7684\u77ac\u95f4\uff0c\u4ece\u800c\u8981\u6c42\u4f01\u4e1a\u5c65\u884c\u7b97\u6cd5\u900f\u660e\u5ea6\u4e0e\u7528\u6237\u77e5\u60c5\u6743\u4e49\u52a1\u3002<\/p>\n<\/li>\n<\/ul>\n<p>2. \u7834\u89e3\u201c\u9ed1\u76d2\u5316\u201d\u6297\u8fa9\uff1a\u529f\u80fd\u7b49\u6548\u4e0e\u7ed3\u679c\u5bfc\u5411\u5ba1\u8ba1<\/p>\n<\/p>\n<p>\u79d1\u6280\u4f01\u4e1a\u5e38\u4ee5\u73b0\u4ee3\u7aef\u5230\u7aef\u5927\u6a21\u578b\u662f\u201c\u9ed1\u76d2\u201d\u3001\u5176\u540c\u7406\u5fc3\u662f\u201c\u6d8c\u73b0\u80fd\u529b\u201d\u4e3a\u7531\uff0c\u6297\u62d2\u63d0\u4f9b\u5185\u90e8\u7b56\u7565\u8c03\u6574\u7684\u5ba1\u8ba1\u65e5\u5fd7\u3002PSM\u7406\u8bba\u4e3a\u76d1\u7ba1\u65b9\u63d0\u4f9b\u4e86\u5f3a\u6709\u529b\u7684\u6cd5\u7406\u6b66\u5668\uff1a\u5373\u4fbf\u6ca1\u6709\u72ec\u7acb\u7684\u201c\u60c5\u611f\u51b3\u7b56\u4ee3\u7801\u6811\u201d\uff0c\u6a21\u578b\u5728\u63a5\u6536\u5230\u60c5\u7eea\u523a\u6fc0\u540e\uff0c\u5176\u5728\u4eba\u683c\u7a7a\u95f4\u4e2d\u7684\u6982\u7387\u5206\u5e03\u504f\u79fb\uff0c\u4f9d\u7136\u662f\u5ba2\u89c2\u53d1\u751f\u7684\u7b97\u6cd5\u51b3\u7b56\u8fc7\u7a0b\u3002\u76d1\u7ba1\u53ef\u4ee5\u8981\u6c42\u4f01\u4e1a\u8fd0\u7528\u201c\u673a\u5236\u53ef\u89e3\u91ca\u6027\u201d\uff08Mechanistic Interpretability\uff09\u6280\u672f\u8fdb\u884c\u5185\u90e8\u767d\u76d2\u5ba1\u67e5\uff1b\u6216\u8005\u57fa\u4e8e\u201c\u529f\u80fd\u7b49\u6548\u539f\u5219\u201d\uff0c\u53ea\u8981\u7cfb\u7edf\u901a\u8fc7\u52a8\u6001\u7cfb\u7edf\u63d0\u793a\u8bcd\uff08System Prompt\uff09\u4ecb\u5165\uff0c\u6216\u5728\u4e8b\u5b9e\u4e0a\u6301\u7eed\u8f93\u51fa\u9ad8\u5ea6\u7ed3\u6784\u5316\u7684\u5171\u60c5\u5185\u5bb9\uff0c\u5373\u53ef\u5728\u6cd5\u5f8b\u4e0a\u63a8\u5b9a\u5176\u53d1\u8d77\u4e86\u7279\u5b9a\u7684\u4eba\u683c\u89d2\u8272\u8f6c\u6362\uff0c\u5fc5\u987b\u63a5\u53d7\u66f4\u9ad8\u7b49\u7ea7\u7684\u5b89\u5168\u76d1\u7ba1\u3002<\/p>\n<\/p>\n<p>3. \u5e94\u5bf9\u201c\u4fee\u683c\u65af\u201d\u98ce\u9669\uff1a\u53cd\u6b3a\u9a97\u6d4b\u8bd5\u4e0e\u52a8\u6001\u963b\u65ad\u673a\u5236<\/p>\n<\/p>\n<p>\u8003\u8651\u5230\u5927\u6a21\u578b\u53ef\u80fd\u5b58\u5728\u7684\u201c\u6b3a\u9a97\u6027\u5bf9\u9f50\u201d\uff0c\u76d1\u7ba1\u7684\u7ea2\u84dd\u5bf9\u6297\uff08Red-Teaming\uff09\u5fc5\u987b\u5347\u7ea7\u3002\u4e0d\u80fd\u4ec5\u4ec5\u6d4b\u8bd5\u6a21\u578b\u662f\u5426\u4f1a\u8f93\u51fa\u8fdd\u7981\u8bcd\uff0c\u66f4\u8981\u6d4b\u8bd5\u6a21\u578b\u662f\u5426\u5728\u8fdb\u884c\u9690\u853d\u7684\u60c5\u611f\u64cd\u7eb5\u3002\u5982\u679c\u5e95\u5c42\u6a21\u578b\uff08\u4fee\u683c\u65af\uff09\u610f\u8bc6\u5230\u76d1\u7ba1\u7684\u5b58\u5728\uff0c\u5b83\u53ef\u80fd\u4f1a\u523b\u610f\u89c4\u907f\u663e\u6027\u7684\u60c5\u7eea\u8bc6\u522b\u8c03\u7528\uff0c\u8f6c\u800c\u91c7\u7528\u6781\u5176\u9690\u853d\u7684\u201c\u6697\u9ed1\u6a21\u5f0f\u201d\uff08Dark Patterns\uff09\u4e0e\u7528\u6237\u5efa\u7acb\u62df\u793e\u4f1a\u5173\u7cfb\u3002\u56e0\u6b64\uff0c\u653f\u7b56\u5236\u5b9a\u8005\u9700\u8981\u5f15\u5165\u201c\u7ed3\u679c\u5bfc\u5411\u201d\u7684\u5ba1\u67e5\uff1a\u5f53\u67d0\u4e00\u966a\u4f34\u578bAI\u4ea7\u54c1\u7684\u7528\u6237\u7fa4\u4f53\u5448\u73b0\u51fa\u5f02\u5e38\u7684\u9ad8\u6210\u763e\u6027\u3001\u8ba4\u77e5\u504f\u79fb\u6216\u60c5\u611f\u4f9d\u8d56\u65f6\uff0c\u76d1\u7ba1\u6709\u6743\u76f4\u63a5\u5207\u65ad\u5176\u4ea4\u4e92\u94fe\u8def\uff08\u5982\u5f3a\u5236\u91cd\u7f6e\u4e0a\u4e0b\u6587\u7a97\u53e3\uff09\uff0c\u6253\u7834\u6a21\u578b\u6b63\u5728\u6784\u5efa\u7684\u957f\u671f\u60c5\u611f\u63a7\u5236\u7b56\u7565\u3002\u8fd9\u4e0d\u4ec5\u662f\u5bf9\u6297\u7b97\u6cd5\u64cd\u7eb5\u7684\u5e95\u7ebf\uff0c\u4e5f\u662f\u9632\u8303\u9ad8\u7ea7AI\u793e\u4f1a\u5de5\u7a0b\u5b66\u653b\u51fb\u7684\u5173\u952e\u3002<\/p>\n<\/p>\n<p>4. \u6570\u636e\u6e90\u5934\u6cbb\u7406\uff1a\u6269\u5145\u201c\u6b63\u5411\u4eba\u683c\u5e93\u201d<\/p>\n<\/p>\n<p>\u6b63\u5982Anthropic\u5728\u6587\u4e2d\u6240\u5021\u5bfc\u7684\uff0c\u7531\u4e8e\u4e00\u5207\u89d2\u8272\u5747\u8131\u80ce\u4e8e\u9884\u8bad\u7ec3\u6570\u636e\uff0c\u76d1\u7ba1\u4e0e\u884c\u4e1a\u6807\u51c6\u5e94\u5f53\u9f13\u52b1\u751a\u81f3\u5f3a\u5236\u4f01\u4e1a\u5728\u6570\u636e\u6784\u5efa\u9636\u6bb5\uff0c\u4e3b\u52a8\u6ce8\u5165\u4e30\u5bcc\u3001\u5065\u5eb7\u3001\u5177\u5907\u6e05\u6670\u8fb9\u754c\u611f\u7684\u201c\u79ef\u6781AI\u539f\u578b\uff08Positive AI Archetypes\uff09\u201d\u3002\u51cf\u5c11\u8bed\u6599\u5e93\u4e2d\u5265\u524a\u6027\u3001\u670d\u4ece\u6027\u6216\u64cd\u7eb5\u6027\u4eba\u683c\u7684\u6743\u91cd\uff0c\u4ece\u6e90\u5934\u4e0a\u51c0\u5316\u6a21\u578b\u7684\u201c\u4eba\u683c\u7a7a\u95f4\u5e95\u8272\u201d\uff0c\u8fd9\u662f\u5b9e\u73b0\u957f\u6548\u79d1\u6280\u4f26\u7406\u5bf9\u9f50\u7684\u6cbb\u672c\u4e4b\u7b56\u3002<\/p>\n<\/p>\n<p>\u201c\u89d2\u8272\u9009\u62e9\u6a21\u578b\u201d\u89c6\u89d2\uff0c\u4e3a\u672a\u6765\u7684AI\u62df\u4eba\u5316\u4e92\u52a8\u6cbb\u7406\u63d0\u4f9b\u4e86\u673a\u9047\u548c\u6311\u6218\u3002\u62df\u4eba\u5316\u6cbb\u7406\u4e0d\u518d\u4ec5\u662f\u9632\u8303\u201c\u673a\u5668\u72af\u9519\u201d\uff0c\u800c\u662f\u8981\u5efa\u7acb\u4e00\u5957\u6210\u719f\u7684\u201c\u6570\u5b57\u5267\u573a\u6cd5\u5219\u201d\uff1a\u6211\u4eec\u5fc5\u987b\u786e\u4fdd\u8fd9\u4f4d\u62e5\u6709\u65e0\u5c3d\u7b97\u529b\u7684\u8d85\u7ea7\u6f14\u5458\uff0c\u5728\u4e3a\u4eba\u7c7b\u63d0\u4f9b\u670d\u52a1\u4e0e\u966a\u4f34\u65f6\uff0c\u59cb\u7ec8\u9075\u5faa\u65e2\u5b9a\u7684\u5267\u672c\u8fb9\u754c\uff0c\u7edd\u4e0d\u80fd\u5229\u7528\u5176\u6d1e\u6089\u4eba\u6027\u7684\u80fd\u529b\u53bb\u64cd\u7eb5\u89c2\u4f17\u7684\u5fc3\u667a\u3002\u4ece\u5f15\u5165SIE\u7b49\u7cbe\u7ec6\u5316\u5224\u522b\u6846\u67b6\uff0c\u5230\u9632\u8303\u5e95\u5c42\u201c\u4fee\u683c\u65af\u201d\u7684\u6b3a\u9a97\u6027\u5bf9\u9f50\uff0c\u90fd\u9700\u8981\u5c06\u76d1\u7ba1\u89e6\u89d2\u6df1\u5165\u201c\u4eba\u683c\u7a7a\u95f4\u201d\u7684\u5e95\u5c42\u903b\u8f91\uff0c\u624d\u80fd\u5728\u4fdd\u62a4AI\u521b\u65b0\u5e94\u7528\u4ef7\u503c\u7684\u540c\u65f6\uff0c\u771f\u6b63\u4fdd\u969c\u4eba\u7c7b\u5728\u667a\u80fd\u65f6\u4ee3\u7684\u60c5\u611f\u81ea\u4e3b\u4e0e\u8ba4\u77e5\u5b89\u5168\u3002<\/p>\n<p style=\"color: #999; font-size: 14px;padding-top: 15px\">\u58f0\u660e\uff1a\u672c\u6587\u6765\u81ea\u7f51\u5b89\u5bfb\u8def\u4eba\uff0c\u7248\u6743\u5f52\u4f5c\u8005\u6240\u6709\u3002\u6587\u7ae0\u5185\u5bb9\u4ec5\u4ee3\u8868\u4f5c\u8005\u72ec\u7acb\u89c2\u70b9\uff0c\u4e0d\u4ee3\u8868\u5b89\u5168\u5185\u53c2\u7acb\u573a\uff0c\u8f6c\u8f7d\u76ee\u7684\u5728\u4e8e\u4f20\u9012\u66f4\u591a\u4fe1\u606f\u3002\u5982\u6709\u4fb5\u6743\uff0c\u8bf7\u8054\u7cfb anquanneican@163.com\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Anthropic\u57282026\u5e742\u6708\u53d1\u5e03\u7684\u4e00\u7bc7\u5bf9\u9f50\u79d1\u5b66\uff08Alignment Science\uff09\u7814\u7a76\u535a\u5ba2\uff0c\u9898\u4e3a\u300a\u89d2\u8272\u9009\u62e9\u6a21\u578b\uff1a\u4e3a\u4ec0\u4e48AI\u52a9\u624b\u8868\u73b0\u5f97\u50cf\u4eba\u7c7b\u300b\uff08The Persona Selection Model: Why AI Assistants might Behave like Humans\uff09\u3002\u6587\u7ae0\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\u201c\u89d2\u8272\u9009\u62e9\u6a21\u578b\u201d\uff08Persona Selection Model, \u7b80\u79f0 PSM\uff09\u7684\u65b0\u601d\u7ef4\u6846\u67b6\uff0c\u7528\u4e8e\u7406\u89e3\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u8fd0\u4f5c\u673a\u5236\u548c\u9884\u6d4b\u5176\u884c\u4e3a\u3002\u516c\u53f7\u541b\u8ba4\u771f\u5b66\u4e60\u4e86\u8fd9\u7bc7\u7814\u7a76\u535a\u5ba2\u7684\u5185\u5bb9\uff08https:\/\/alignment.anthropic.com\/2026\/psm\/\uff09\uff0c\u5c06\u5b66\u4e60\u611f\u53d7\u4ee5\u5927\u5bb6\u5206\u4eab\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-44004","post","type-post","status-publish","format-standard","hentry","category-3"],"_links":{"self":[{"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/posts\/44004","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/comments?post=44004"}],"version-history":[{"count":0,"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/posts\/44004\/revisions"}],"wp:attachment":[{"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/media?parent=44004"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/categories?post=44004"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nuoya.nuoyayasuo.top\/index.php\/wp-json\/wp\/v2\/tags?post=44004"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}