{"id":12621,"date":"2025-01-23T15:24:30","date_gmt":"2025-01-23T07:24:30","guid":{"rendered":"https:\/\/ihower.tw\/blog\/?p=12621"},"modified":"2025-07-04T07:42:52","modified_gmt":"2025-07-03T23:42:52","slug":"aie-common-pitfalls-ai-app","status":"publish","type":"post","link":"https:\/\/ihower.tw\/blog\/12621-aie-common-pitfalls-ai-app","title":{"rendered":"\u611b\u597d AI Engineer \u96fb\u5b50\u5831 \ud83d\ude80 AI \u61c9\u7528\u958b\u767c\u7684\u5e38\u898b\u932f\u8aa4 #22"},"content":{"rendered":"\n
\n

\u6b61\u8fce\u8a02\u95b1 \ud83d\udcec<\/a> \u611b\u597d AI Engineer \u96fb\u5b50\u5831 \u904e\u5f80\u671f\u6578\u9ede\u9019<\/a> \ud83d\udcda<\/p>\n<\/blockquote>\n\n\n\n

\"cover\"\/<\/figure>\n\n\n\n

Hello! \u5404\u4f4d AI \u958b\u767c\u8005\u5927\u5bb6\u597d \ud83d\udc4b<\/p>\n\n\n\n

\u6211\u662f ihower\uff0c\u8fb2\u66c6\u65b0\u5e74\u7684\u8173\u6b65\u63a5\u8fd1\uff0c\u5148\u5728\u9019\u88e1\u795d\u798f\u5927\u5bb6\u65b0\u5e74\u5feb\u6a02\uff0c\u842c\u4e8b\u5982\u610f\uff0c\u8ca1\u6e90\u6efe\u6efe\uff01 \ud83e\udde7\ud83e\udde8<\/p>\n\n\n\n

\ud83d\udd1d Common pitfalls when building generative AI applications<\/a><\/h3>\n\n\n\n

\u9019\u7bc7 Chip Huyen \u5beb\u7684 “\u751f\u6210\u5f0f AI \u61c9\u7528\u958b\u767c\u7684\u5e38\u898b\u932f\u8aa4”\uff0c\u6b78\u7d0d\u51fa 6 \u500b\u5bb9\u6613\u8e29\u96f7\u7684\u5730\u65b9\uff0c\u5be6\u5728\u5fc3\u6709\u621a\u621a\u7109\u5566\u3002\u4ee5\u4e0b\u7d50\u5408\u539f\u6587\u8207\u6211\u81ea\u5df1\u7684\u7d93\u9a57\uff0c\u6574\u7406\u51fa\u4ee5\u4e0b\u516d\u500b\u89c0\u9ede\uff1a<\/p>\n\n\n\n

1\ufe0f\u20e3 \u4e0d\u8a72\u7528\u751f\u6210\u5f0f AI \u7684\u5834\u666f\u786c\u8981\u7528\u751f\u6210\u5f0f AI<\/p>\n\n\n\n

\u5927\u6a21\u578b LLM \u662f\u5f88\u53b2\u5bb3\uff0c\u4f46\u662f\u5f88\u591a\u4eba\u5c0d\u65bc\u9019\u9805\u6280\u8853\u7684\u80fd\u529b\u4e0d\u5920\u4e86\u89e3\uff0c\u786c\u8981\u628a\u4e0d\u9069\u5408\u7684\u9700\u6c42\u585e\u7d66\u5b83\u8655\u7406\u3002\u4f8b\u5982: \u6700\u4f73\u5316\u80fd\u6e90\u6d88\u8017\u3001\u7528\u4f86\u6aa2\u6e2c\u6d41\u91cf\u7570\u5e38\u3001\u7528\u4f86\u9810\u6e2c\u96fb\u91cf\u3001\u6aa2\u6e2c\u75c5\u4eba\u662f\u5426\u71df\u990a\u4e0d\u826f\u7b49\u7b49\uff0c\u5176\u5be6\u90fd\u4e0d\u9069\u5408\u7528\u751f\u6210\u5f0f AI \u4f86\u505a\u3002<\/p>\n\n\n\n

\u6211\u500b\u4eba\u4e5f\u78b0\u904e\u516c\u53f8\u60f3\u8981\u505a\u4e9b\u71df\u904b\u6700\u4f73\u5316\u548c\u7570\u5e38\u6aa2\u6e2c\uff0c\u7d30\u554f\u624d\u767c\u73fe\u8cc7\u6599\u90fd\u662f\u6578\u5b57\u578b\u614b\u3002\u9019\u7a2e\u4e0d\u662f\u7528\u751f\u6210\u5f0f AI \u5566\uff0c\u800c\u662f\u61c9\u8a72\u7528\u5c08\u9580\u7684\u6f14\u7b97\u6cd5\u6216\u662f\u6a5f\u5668\u5b78\u7fd2\u4f86\u505a\u3002<\/p>\n\n\n\n

2\ufe0f\u20e3 \u641e\u6df7\u300c\u721b AI \u7522\u54c1\u300d\u8207\u300c\u7b28 AI \u6a21\u578b\u300d<\/p>\n\n\n\n

\u5f88\u591a\u5931\u6557\u6848\u4f8b\u5176\u5be6\u4e0d\u662f LLM \u6a21\u578b\u4e0d\u5920\u8070\u660e\uff0c\u800c\u662f\u7522\u54c1\u8a2d\u8a08\u548c AI \u5de5\u7a0b\u80fd\u529b\u7684\u4e0d\u8db3\u3002<\/p>\n\n\n\n

\u4f8b\u5982\u6700\u8fd1\u6700\u706b\u71b1\u7684 AI Coding \u7522\u54c1\u70ba\u4f8b\uff0c\u7121\u8ad6\u662f GitHub Copilot\u3001Cursor\u3001Windsurf \u9084\u662f Devin\uff0c\u80cc\u5f8c\u7528\u7684\u5f80\u5f80\u90fd\u662f\u4e00\u6a23\u7684 LLM\uff0c\u5dee\u5225\u5728\u65bc\u7522\u54c1\u8a2d\u8a08\u8207 AI \u5de5\u7a0b\u505a\u5f97\u597d\u4e0d\u597d\u3002\u6240\u8b02\u7684 AI \u5de5\u7a0b\u80fd\u529b\uff0c\u4e3b\u8981\u662f\u6307\u5982\u4f55\u6709\u6548\u63d0\u4f9b\u6a21\u578b\u6240\u9700\u7684 context\uff08\u900f\u904e prompt chaining\u3001RAG \u7b49\u6280\u8853)\uff0c\u8b93\u6a21\u578b\u80fd\u5728\u6b63\u78ba\u8cc7\u6599\u7684\u57fa\u790e\u4e0a\u7522\u751f\u6709\u7528\u7684\u7d50\u679c\u3002\u82e5\u7f3a\u4e4f\u9019\u4e9b\u5de5\u7a0b\u74b0\u7bc0\uff0c\u518d\u5f37\u5927\u7684\u6a21\u578b\u4e5f\u53ea\u80fd”\u5de7\u5a66\u96e3\u70ba\u7121\u7c73\u4e4b\u708a”<\/p>\n\n\n\n

\u56e0\u6b64\u4f60\u6703\u770b\u5230\u51fa\u73fe o3 \u9019\u7a2e\u53ef\u4ee5\u89e3\u6c7a\u5967\u6578\u7684\u8070\u660e\u6a21\u578b\uff0c\u53c8\u540c\u6642\u770b\u5230 Devin \u8655\u7406\u5be6\u969b\u7684\u8edf\u9ad4\u958b\u767c\u4efb\u52d9\u9084\u505a\u4e0d\u597d\u3002\u6839\u672c\u539f\u56e0\u53ef\u80fd\u4e0d\u662f\u6a21\u578b\u592a\u5f31\uff0c\u800c\u662f\u6574\u5408\u7684 AI \u5de5\u7a0b\u9084\u4e0d\u5920\u5b8c\u5584\u3002<\/p>\n\n\n\n

\u6700\u8fd1\u770b\u5230\u4e00\u53e5\u8a71\u662f raw intelligence \u2260 intelligent software systems<\/strong>\u300c\u539f\u59cb\u667a\u529b\u300d\u4e0d\u7b49\u65bc\u300c\u667a\u6167\u8edf\u9ad4\u7cfb\u7d71\u300d<\/p>\n\n\n\n

\u5927\u6a21\u578b\u7684\u300c\u667a\u529b\u300d\u53ea\u662f\u57fa\u77f3\uff0c\u8981\u628a\u5b83\u8f49\u5316\u6210\u771f\u6b63\u6709\u6548\u7684\u667a\u6167\u7cfb\u7d71\uff0c\u9084\u9700\u8981\u6b63\u78ba\u7684\u4e0a\u4e0b\u6587\u3001\u5de5\u5177\u6574\u5408\u4ee5\u53ca\u6210\u719f\u7684\u5de5\u7a0b\u5316\u6d41\u7a0b\uff0c\u9019\u5c31\u662f Berkeley AI Research (BAIR) \u6240\u63d0\u51fa\u7684 The Shift from Models to Compound AI Systems<\/a> \u7684\u8da8\u52e2\uff1a\u9818\u5148\u7684 AI \u6210\u679c\u8d8a\u4f86\u8d8a\u4f9d\u8cf4\u300c\u591a\u91cd\u5143\u4ef6\u300d\u5354\u4f5c\uff0c\u800c\u975e\u55ae\u7d14\u4f9d\u8cf4\u4e00\u500b\u5927\u578b\u6a21\u578b\uff0c\u624d\u80fd\u767c\u63ee LLM \u7684\u6700\u5927\u6f5b\u80fd\uff0c\u505a\u51fa\u597d\u7528\u7684 AI \u7522\u54c1\u3002<\/p>\n\n\n\n\n\n\n\n

3\ufe0f\u20e3 \u4e00\u958b\u59cb\u5c31\u641e\u5f97\u592a\u8907\u96dc<\/p>\n\n\n\n

\u5e38\u898b\u7684\u904e\u65e9\u8907\u96dc\u5316\u5305\u62ec 1. \u4e0d\u9700\u8981\u6642\u5c31\u7528 agent \u6846\u67b6 2. \u7cfe\u7d50\u9078\u64c7\u5411\u91cf\u8cc7\u6599\u5eab 3. \u660e\u660e prompting \u5c31\u8db3\u5920\u9084\u8981\u641e fine-tuning \u7b49\u7b49\u3002\u592a\u65e9\u5f15\u5165\u8907\u96dc\u7684\u5de5\u5177\u6703\u62bd\u8c61\u6389\u95dc\u9375\u7d30\u7bc0\uff0c\u8b93\u7cfb\u7d71\u66f4\u96e3\u7406\u89e3\u548c\u9664\u932f\u3002<\/p>\n\n\n\n

4\ufe0f\u20e3 \u904e\u5ea6\u6a02\u89c0\u770b\u5f85\u521d\u671f\u6210\u529f<\/p>\n\n\n\n

\u6211\u4e4b\u524d\u4e5f\u5206\u4eab\u904e\u7684 LinkedIn \u958b\u767c\u7d93\u9a57<\/a>: \u5f9e 0 \u5230 80% \u53ea\u82b1\u4e00\u500b\u6708\uff0c\u4f46\u5f9e 80% \u5230 95% \u537b\u82b1\u4e86\u56db\u500b\u6708\uff0c\u6700\u5f8c\u90a3 10-20% \u7684\u512a\u5316\u5f80\u5f80\u6700\u8017\u6642\u96e3\u5ea6\u6700\u9ad8\u3002\u505a\u51fa\u4e00\u500b Demo \u5f88\u5bb9\u6613\uff0c\u4f46\u8981\u505a\u51fa\u4e00\u500b\u771f\u6b63\u80fd\u4e0a\u7dda\u7684\u7522\u54c1\u537b\u5f88\u56f0\u96e3: \u5e7b\u89ba\u3001\u5ef6\u9072\u3001\u6e96\u78ba\u5ea6\u8207\u6548\u80fd\u7684\u6b0a\u8861\u3001Prompt \u8a2d\u8a08\u3001\u8a55\u4f30\u7b49\u7b49\u7b49\uff0c\u9019\u4e9b\u5728\u958b\u767c\u5f8c\u671f\u90fd\u662f\u5927\u8003\u9a57\u3002<\/p>\n\n\n\n

5\ufe0f\u20e3 \u5ffd\u8996\u4eba\u5de5\u8a55\u4f30<\/p>\n\n\n\n

\u96d6\u7136\u7528 AI \u4f86\u8a55\u4f30 AI (LLM-as-judge)\u5f88\u65b9\u4fbf\uff0c\u4f46\u4e0d\u80fd\u5b8c\u5168\u4f9d\u8cf4\u3002\u6700\u597d\u7684\u5718\u968a\u90fd\u6703\u6bcf\u5929\u62bd\u6a23\u505a\u4eba\u5de5\u8a55\u4f30(30-1000 \u4f8b)\uff0c\u7528\u4f86\u88dc\u5145\u4ed6\u5011\u7684\u81ea\u52d5\u5316\u8a55\u4f30\uff0c \u4e26\u4e14\u53ef\u4ee5 1. \u9a57\u8b49 AI \u8a55\u4f30\u7684\u6e96\u78ba\u6027 2. \u4e86\u89e3\u7528\u6236\u4f7f\u7528\u60c5\u6cc1 3. \u53ca\u65e9\u767c\u73fe\u7570\u5e38\u6a21\u5f0f<\/p>\n\n\n\n

6\ufe0f\u20e3 \u76f2\u76ee\u8490\u96c6\u4f7f\u7528\u6848\u4f8b<\/p>\n\n\n\n

\u4e0d\u5c11\u4f01\u696d\u60f3\u505a AI \u6578\u4f4d\u8f49\u578b\uff0c\u4f46\u5c0d\u65bc\u76ee\u6a19\u4e0d\u5920\u660e\u78ba(\u5c31\u662f\u4e00\u7a2e\u6211\u77e5\u9053 LLM \u5f88\u53b2\u5bb3\uff0c\u4f46\u6211\u4e0d\u77e5\u9053\u53ef\u4ee5\u7528\u5728\u54ea\u88e1)\uff0c\u65bc\u662f\u5c31\u8b93\u54e1\u5de5\u5929\u99ac\u884c\u7a7a\u5730\u63d0\u6848\u4f86\u500b\u9ed1\u5ba2\u677e\uff0c\u641e\u51fa\u4e00\u5927\u5806\u91cd\u8907\u7684\u5c0f\u5c08\u6848(\u50cf\u662f\u5230\u8655\u90fd\u5728\u505a\u804a\u5929\u6a5f\u5668\u4eba\u6216\u662f\u5404\u7a2e\u63d2\u4ef6)\u3002\u7136\u5f8c AI \u958b\u767c\u5718\u968a\u5c31\u88ab\u9019\u4e9b\u96f6\u788e\u3001\u5f71\u97ff\u529b\u6709\u9650\u7684\u5c08\u6848\u5206\u6563\u6ce8\u610f\u529b\uff0c\u6700\u5f8c\u516c\u53f8\u5f97\u51fa “\u751f\u6210\u5f0f AI \u6295\u8cc7\u5831\u916c\u7387\u4e0d\u9ad8” \u7684\u7d50\u8ad6\u3002\u9019\u88e1\u5efa\u8b70\u61c9\u8a72\u8981\u6709\u5927\u5c40\u89c0\u548c\u6574\u9ad4\u7b56\u7565\uff0c\u5c08\u6ce8\u5728\u771f\u6b63\u80fd\u5e36\u4f86\u9ad8\u50f9\u503c\u7684\u61c9\u7528\u4e0a\uff0c\u9019\u6a23\u624d\u80fd\u6709\u6700\u5927\u7684\u56de\u5831\u3002<\/p>\n\n\n\n

\u4ee5\u4e0a\uff0c\u5c0d\u65bc\u6b63\u5728\u898f\u5283\u6216\u958b\u767c AI \u61c9\u7528\u7684\u5718\u968a\u4f86\u8aaa\uff0c\u9019\u4e9b\u5751\u90fd\u883b\u5e38\u898b\u7684\uff0c\u4e0d\u53ea\u662f\u6280\u8853\u554f\u984c\uff0c\u4e5f\u9700\u8981\u6709\u6e05\u695a\u7684\u7b56\u7565\u548c\u52d9\u5be6\u7684\u5de5\u7a0b\u601d\u7dad\u554a\u3002<\/p>\n\n\n\n

\u66f4\u591a\u8a0e\u8ad6\u5728\u6211 Facebook<\/a> \u8cbc\u6587\u3002<\/p>\n\n\n\n

\ud83e\udde0 o1 isn\u2019t a chat model<\/a><\/h3>\n\n\n\n

“OpenAI o1 \u4e0d\u662f\u4e00\u500b\u804a\u5929\u7528\u7684\u6a21\u578b\uff0c\u800c\u662f\u4e00\u500b\u5831\u544a\u7522\u751f\u5668!” \u6700\u8fd1\u770b\u5230\u9019\u7a2e\u8aaa\u6cd5\u89ba\u5f97\u5f88\u8cbc\u5207\uff0c\u4e5f\u7b26\u5408\u6211\u6700\u8fd1\u7684\u4f7f\u7528\u9ad4\u9a57 \ud83d\ude00<\/p>\n\n\n\n

ChatGPT \u53bb\u5e74\u5e95\u63a8\u51fa\u7684\u63a8\u7406\u578b o1 \u6a21\u578b\u662f\u4e00\u7a2e\u4e0d\u540c\u985e\u578b\u7684\u6a21\u578b\uff0c\u5b83\u6703\u5148\u5167\u90e8\u601d\u8003\u5f8c\u624d\u56de\u7b54\u4f60\u3002\u4f60\u9700\u8981\u7528\u4e0d\u540c\u7684 prompting \u65b9\u5f0f\u4f86\u4f7f\u7528\u5b83\u3002<\/p>\n\n\n\n

\u6700\u8fd1\u770b\u5230\u8d8a\u4f86\u8d8a\u591a\u4eba\u5728\u8a0e\u8ad6\u5206\u4eab\u5982\u4f55\u91dd\u5c0d o1 \u6a21\u578b\u4e0b prompt\uff0c\u6709\u770b\u5230\u4e00\u7bc7\u4e0d\u932f\u7684\u7d93\u9a57\u5206\u4eab\u6587\u7ae0(\u9023\u7d50\u5728\u7559\u8a00)\uff0c\u6574\u7406\u4e00\u4e9b\u91cd\u9ede\u5982\u4e0b:<\/p>\n\n\n\n

    \n
  1. \u9019\u4e0d\u662f\u4e00\u500b\u804a\u5929\u6a21\u578b\uff0c\u800c\u662f\u4e00\u500b\u5831\u544a\u7522\u751f\u5668!<\/li>\n\n\n\n
  2. \u76e1\u53ef\u80fd\u63d0\u4f9b\u66f4\u591a\u53c3\u8003\u8cc7\u6599(Context)\uff0c\u4e26\u6e05\u695a\u8aaa\u660e\u60a8\u7684\u9700\u6c42\u548c\u76ee\u6a19(Goal)<\/li>\n\n\n\n
  3. \u4e0d\u9700\u8981\u82b1\u6642\u9593\u96d5\u7422 prompt \u63d0\u793a\u8a5e\uff0c\u800c\u662f\u8981\u585e\u66f4\u591a\u554f\u984c\u63cf\u8ff0\u548c\u5927\u91cf\u80cc\u666f\u8cc7\u8a0a<\/li>\n\n\n\n
  4. \u4e0d\u8981\u8ddf\u5b83\u8aaa “how”\uff0c\u53ea\u8981\u8aaa “what”\uff0c\u5c08\u6ce8\u5728\u7d66\u76ee\u6a19\u800c\u4e0d\u662f\u7d66\u65b9\u6cd5<\/li>\n\n\n\n
  5. prompt \u53ef\u4ee5\u5305\u62ec criteria \u5982\u4f55\u5224\u65b7\u8f38\u51fa\u597d\u8ddf\u58de\u7684\u6a19\u6e96\uff0c\u9019\u6a23 o1 \u53ef\u4ee5\u66f4\u597d\u81ea\u6211\u6539\u9032\u3001\u4fee\u6b63\u932f\u8aa4<\/li>\n\n\n\n
  6. o1 \u64c5\u9577\u751f\u6210\u5b8c\u6574\u5831\u544a\u548c\u5206\u6790\u3001\u89e3\u91cb\u8907\u96dc\u6982\u5ff5\u3001\u4e00\u6b21\u8655\u7406\u5927\u91cf\u5167\u5bb9<\/li>\n\n\n\n
  7. o1 \u4e0d\u64c5\u9577\u7528\u7279\u5b9a\u7684\u5beb\u4f5c\u98a8\u683c<\/li>\n\n\n\n
  8. \u8ddf o1 \u4e92\u52d5\u66f4\u50cf\u662f\u5beb email \u6c42\u52a9\u800c\u4e0d\u662f\u804a\u5929: \u4f60\u6703\u4e00\u6b21\u7d66\u6bd4\u8f03\u5b8c\u6574\u7684\u80cc\u666f\u8cc7\u6599\u548c\u6240\u6709\u8cc7\u8a0a\u7d66\u5c0d\u65b9<\/li>\n<\/ol>\n\n\n\n

    \u4ee5\u4e0b\u622a\u5716\u662f\u6587\u7ae0\u4e2d\u7684 o1 prompt \u7bc4\u4f8b:<\/p>\n\n\n\n

    \"\"\/<\/figure>\n\n\n\n

    \u6ce8\u610f: \u4ee5\u4e0a tips \u5c0d\u65bc\u4e00\u822c\u6a21\u578b\u4f8b\u5982 gpt-4o, claude, gemini \u7b49\u4e0d\u9069\u7528\u5594<\/p>\n\n\n\n

    \u66f4\u591a\u8a0e\u8ad6\u5728\u6211 Facebook<\/a> \u8cbc\u6587\u3002<\/p>\n\n\n\n

    \ud83d\udc4d The 6 AI Engineering Patterns In 2025<\/a><\/h3>\n\n\n\n

    \u7531 Greg Kamradt (\u73fe\u70ba ARC Prize \u7684 President) \u64b0\u5beb\u7684\u9019\u7bc7 “The 6 AI Engineering Patterns In 2025” \u8b1b\u4e86 AI \u5de5\u7a0b\u5305\u62ec\u54ea\u4e9b\u6771\u897f\uff0c\u4ee5\u53ca\u4e00\u4e9b\u5b78\u7fd2\u8cc7\u6e90:<\/p>\n\n\n\n

    1. Models (\u6a21\u578b)<\/p>\n\n\n\n

      \n
    1. <\/li>\n<\/ol>\n\n\n\n
        \n
      1. <\/li>\n<\/ol>\n\n\n\n