\u5168\u6587\u7ffb\u8bd1\u7248\u672c<\/a>\uff09\uff0c\u5bf9\u4e8e Gemini \u7684\u8bad\u7ec3\u539f\u7406\u548c\u6d4b\u8bd5\u7b49\u8fdb\u884c\u4e86\u8be6\u7ec6\u89e3\u91ca\u3002<\/span><\/p>\n\u5230\u5e95\u5982\u4f55\u7406\u89e3 Gemini \u7684\u539f\u751f\u591a\u6a21\u6001\u80fd\u529b\uff1f<\/span><\/p>\n\u4ee5\u53ca\u4ece\u73b0\u5728\u7684\u62a5\u544a\u6570\u636e\u6765\u770b\uff0cGemini \u7684\u80fd\u529b\u548c GPT-4 \u76f8\u6bd4\uff0c\u5dee\u8ddd\u6709\u591a\u5927\uff1f<\/span><\/p>\n\u6211\u4eec\u9009\u53d6\u4e86\u4e24\u4f4d\u76f8\u5173\u884c\u4e1a\u6280\u672f\u4e13\u5bb6\u5bf9\u4e8e\u8fd9\u4efd\u6280\u672f\u62a5\u544a\u3001Gemini \u4ea7\u54c1\u6f14\u793a\u89c6\u9891\u7684\u5206\u6790\uff0c\u5728 Gemini Ultra \u672a\u53d1\u5e03\u4e4b\u524d\uff0c\u63d0\u4f9b\u66f4\u591a\u89d2\u5ea6\u7684\u89e3\u8bfb\u3002<\/span><\/p>\nFounder Park \u5df2\u83b7\u53d6\u4e24\u4f4d\u4f5c\u8005\u6388\u6743\u8f6c\u8f7d\u3002<\/span><\/p>\n<\/p>\n
<\/h2>\n\u674e\u535a\u6770<\/strong><\/span><\/h2>\nGemini \u771f\u5b9e\u80fd\u529b\u76f8\u6bd4 GPT-4 \u5e94\u8be5\u6709\u4e00\u5b9a\u5dee\u8ddd<\/strong><\/span><\/h2>\n\u73b0\u5728\u8fde Google \u90fd\u5f00\u59cb\u5237\u699c\u4e86<\/span><\/strong><\/span><\/h3>\n\u521a\u521a\u8ddf\u6211\u4eec co-founder \u8ba8\u8bba\u4e86\u4e0b\uff0c\u4ed6\u662f evaluation \u7684\u8001\u624b\uff0c\u5370\u8bc1\u4e86\u6211\u7684\u731c\u6d4b\u3002<\/span><\/p>\n\u9996\u5148\u8ddf\u00a0GPT-4\u00a0\u5bf9\u6bd4\u7684\u65f6\u5019\uff0c\u7adf\u7136\u662f\u81ea\u5df1\u7528 CoT\uff0cGPT-4 \u7528\u00a0few-shot\uff0c\u8fd9\u672c\u8eab\u5c31\u4e0d\u516c\u5e73<\/strong>\u3002CoT\uff08\u601d\u7ef4\u94fe\uff09\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002\u6709\u6ca1\u6709 CoT \u7684\u533a\u522b\uff0c\u5c31\u597d\u50cf\u8003\u8bd5\u7684\u65f6\u5019\u4e00\u4e2a\u4eba\u5141\u8bb8\u7528\u8349\u7a3f\u7eb8\uff0c\u53e6\u4e00\u4e2a\u4eba\u53ea\u5141\u8bb8\u53e3\u7b97\u3002<\/span><\/p>\n\u66f4\u5938\u5f20\u7684\u662f\uff0c\u7528\u4e86 CoT@32\uff0c\u4e5f\u5c31\u662f\u6bcf\u4e2a\u95ee\u9898\u56de\u7b54 32 \u6b21\uff0c\u9009\u51fa\u5176\u4e2d\u51fa\u73b0\u6b21\u6570\u6700\u591a\u7684\u90a3\u4e2a\u7b54\u6848\u4f5c\u4e3a\u8f93\u51fa\u3002\u4e5f\u5c31\u662f\u8bf4\u660e Gemini \u7684\u5e7b\u89c9\u5f88\u4e25\u91cd\uff0c\u540c\u4e00\u4e2a\u95ee\u9898\u56de\u7b54\u51c6\u786e\u7387\u4e0d\u9ad8\uff0c\u6240\u4ee5\u624d\u9700\u8981\u91cd\u590d\u56de\u7b54 32 \u6b21\u9009\u51fa\u73b0\u6b21\u6570\u6700\u591a\u7684\u3002\u751f\u4ea7\u73af\u5883\u4e2d\u771f\u8981\u8fd9\u4e48\u641e\uff0c\u6210\u672c\u5f97\u591a\u9ad8\u5440\uff01<\/span><\/p>\n<\/p>\n
\u5176\u6b21\u662f\u7528\u4e86\u672a\u5bf9\u9f50\u7684\u6a21\u578b\u8ddf\u5df2\u7ecf\u5bf9\u9f50\u7684 GPT-4 \u505a\u5bf9\u6bd4\u3002<\/span><\/strong>GPT-4 \u62a5\u544a\u91cc\u9762\u5df2\u7ecf\u8bf4\u4e86\u6a21\u578b\u5bf9\u9f50\u4f1a\u964d\u4f4e\u77e5\u8bc6\u65b9\u9762\u7684\u80fd\u529b\uff0c\u4f46\u662f\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002\u4e4b\u524d\u6211\u4eec\u7528 GPT-3.5 \u672a\u5bf9\u9f50\u7684\u5185\u90e8\u7248\u672c\u505a\u6d4b\u8bd5\uff0c\u53d1\u73b0\u53ef\u4ee5\u77e5\u9053\u4e2d\u79d1\u5927\u67d0\u4e2a\u6559\u6388\u6559\u4e86\u54ea\u95e8\u8bfe\u8fd9\u79cd\u7ea7\u522b\u7684\u7ec6\u8282\uff0c\u4f46\u516c\u5f00\u53d1\u5e03\u7684\u5df2\u5bf9\u9f50\u7248\u672c\u5c31\u53ea\u80fd\u77e5\u9053\u4e2d\u79d1\u5927\u7684\u6821\u957f\u662f\u8c01\u4e86\u3002\u6240\u4ee5\u7528\u672a\u5bf9\u9f50\u7684 Gemini \u548c\u5df2\u5bf9\u9f50\u7684 GPT-4 \u5bf9\u6bd4\u4e5f\u662f\u4e0d\u592a\u516c\u5e73\u7684\u3002<\/span><\/p>\nGemini \u7684\u771f\u5b9e\u80fd\u529b\u80af\u5b9a\u662f\u8fdc\u8d85 GPT-3.5 \u7684\uff0c\u80af\u5b9a\u662f\u4e2a\u6bd4\u8f83\u9760\u8c31\u7684\u6a21\u578b\uff0c\u4f46\u662f\u76f8\u6bd4 GPT-4 \u4f30\u8ba1\u8fd8\u6709\u5dee\u8ddd\u3002<\/span><\/p>\n\u76ee\u524d Gemini \u8fd8\u6ca1\u6709\u516c\u5e03\u5b9a\u4ef7\uff0c\u5982\u679c Ultra \u6a21\u578b\u7684\u5b9a\u4ef7\u662f GPT-3.5 \u7684\u91cf\u7ea7\uff0c\u90a3\u4e48\u5b83\u7684\u80fd\u529b\u663e\u7136\u662f\u6bd4 GPT-3.5 \u66f4\u5f3a\u7684\uff0c\u503c\u5f97\u7528\u3002\u4f46\u662f\u5982\u679c\u5b9a\u4ef7\u7c7b\u4f3c GPT-4\uff0c\u90a3\u4e48\u53ef\u80fd\u8fd8\u662f GPT-4 \u66f4\u5b9e\u7528\u4e00\u4e9b\u3002<\/span><\/p>\n\u89c6\u9891\u7406\u89e3\uff1a\u5b9e\u65f6\u6027\u6bd4\u51c6\u786e\u6027\u66f4\u91cd\u8981<\/span><\/strong><\/span><\/h3>\nGemini \u89c6\u9891\u7406\u89e3\u7684\u80fd\u529b\u4e0d\u9519<\/span><\/strong>\uff0c\u6f14\u793a\u89c6\u9891\u5f88\u9177\u70ab\u3002\u4f46\u53ef\u60dc\u7684\u662f\u8fd9\u4e2a\u89c6\u9891\u662f\u526a\u8f91\u51fa\u6765\u7684\uff0c\u5b9e\u9645\u7684 Gemini \u6839\u672c\u8fbe\u4e0d\u5230\u6f14\u793a\u89c6\u9891\u7684\u5b9e\u65f6\u6027\u3002<\/strong><\/span><\/p>\n\u5176\u5b9e\u8fd9\u4e2a\u6f14\u793a\u89c6\u9891\u4e2d\u7684\u6548\u679c GPT-4V \u4e5f\u80fd\u505a\u51fa\u6765\uff0c\u53ea\u8981\u628a\u622a\u56fe\u5582\u7ed9 GPT-4V \u8fd9\u4e9b\u4efb\u52a1\u5c31\u90fd\u80fd\u5b8c\u6210\u4e86\uff08\u751f\u6210\u56fe\u7247\u7684\u4efb\u52a1\u53ef\u4ee5\u8f6c\u6210\u6587\u672c\u518d\u63a5\u56fe\u7247\u751f\u6210\u6a21\u578b\uff09\uff0c\u5f53\u7136 GPT-4V \u7684\u5ef6\u8fdf\u6bd4\u8f83\u9ad8\uff0c\u505a\u4e0d\u5230 Gemini \u8fd9\u4e2a\u89c6\u9891\u91cc\u8fd9\u4e48\u5b9e\u65f6\u3002\u6211\u8fd8\u6ca1\u6709\u7528\u8fc7 Gemini \u7684 API\uff0c\u4e0d\u77e5\u9053\u5b9e\u9645\u5ef6\u8fdf\u4f1a\u4e0d\u4f1a\u6bd4 GPT-4V \u66f4\u4f4e\u3002<\/span><\/p>\n\u4e00\u4e9b\u6bd4\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u6bd4\u5982 Fuyu-8B \u548c MiniGPT-v2\uff0c\u4e5f\u80fd\u505a\u5230\u8fd9\u4e2a\u6f14\u793a\u89c6\u9891\u4e2d\u5927\u90e8\u5206\u7684\u6548\u679c\uff0c\u8fd9\u4e9b\u4efb\u52a1\u90fd\u662f VQA \u91cc\u9762\u76f8\u5bf9\u57fa\u672c\u7684\u3002\u8fd9\u4e9b\u5c0f\u7684\u5f00\u6e90\u6a21\u578b\u8fd8\u6709\u4e2a\u4f18\u52bf\uff0c\u4ece\u56fe\u7247\u8f93\u5165\u5230\u9996\u4e2a token \u8f93\u51fa\u7684\u5ef6\u8fdf\u53ea\u6709 100-200 ms\uff0c\u53ef\u4ee5\u505a\u5230\u8fd9\u4e2a\u6f14\u793a\u89c6\u9891\u91cc\u9762\u7684\u5b9e\u65f6\u6548\u679c\u3002\u56fe\u7247\u751f\u6210\u65b9\u9762\uff0cstable diffusion 20 \u4e2a step \u80af\u5b9a\u8fbe\u4e0d\u5230\u8fd9\u4e2a\u6f14\u793a\u89c6\u9891\u91cc\u7684\u65f6\u5ef6\uff0c\u9700\u8981\u7528 SDXL Turbo \u6216\u8005 LCM \u8fd9\u4e9b\u6700\u65b0\u7684\u6a21\u578b\u624d\u80fd\u505a\u5230\u3002<\/span><\/p>\n\u4ece\u7528\u6237\u4f53\u9a8c\u89d2\u5ea6\u770b\uff0ctoC \u573a\u666f\u4e0b\u5b9e\u65f6\u6027\u5176\u5b9e\u662f\u6bd4\u51c6\u786e\u7387\u66f4\u91cd\u8981\u7684<\/span><\/strong>\u3002\u6bd4\u5982\u8bed\u97f3\u8bc6\u522b\uff0c\u867d\u7136 OpenAI \u7684 Whisper API \u51c6\u786e\u7387\u662f\u5f88\u9ad8\u7684\uff0cVITS \u5408\u6210\u7684\u58f0\u97f3\u4e5f\u6bd4\u8f83\u81ea\u7136\uff0c\u4f46\u662f\u8bed\u97f3\u8bc6\u522b\u548c\u8bed\u97f3\u5408\u6210\u76ee\u524d\u90fd\u662f\u4ee5\u6574\u53e5\u4e3a\u5355\u4f4d\u8fdb\u884c\u7684\uff0c\u5373\u4f7f\u6309\u7167\u53e5\u5b50\u5207\u7247\u505a\u8bc6\u522b\u548c\u5408\u6210\uff0c\u4e00\u4e2a\u8bed\u97f3\u5bf9\u8bdd\u7cfb\u7edf\u7684\u7aef\u5230\u7aef\u65f6\u5ef6\uff08\u4ece\u7528\u6237\u8bf4\u8bdd\u7ed3\u675f\u5230 AI \u5f00\u59cb\u8bf4\u8bdd\uff09\u9ad8\u8fbe 5 \u79d2\u5de6\u53f3\uff0c\u8fd9\u662f\u7528\u6237\u96be\u4ee5\u5fcd\u53d7\u7684\u3002Whisper \u548c VITS \u539f\u751f\u90fd\u4e0d\u652f\u6301\u7c7b\u4f3c\u540c\u58f0\u4f20\u8bd1\u7684 streaming\u3002\u8981\u60f3\u505a\u5230 2 \u79d2\u4ee5\u5185\u7684\u8bed\u97f3\u5ef6\u8fdf\uff0c\u8fd8\u662f\u9700\u8981\u5f88\u591a\u5de5\u7a0b\u4f18\u5316\u7684\u3002<\/span><\/p>\n\u4ee5\u540e\u53d1\u5e03\u7684\u652f\u6301\u8bed\u97f3\u901a\u8bdd\u7684\u4ea7\u54c1\uff0c\u53ef\u4ee5\u62ff Gemini \u8fd9\u4e2a\u6f14\u793a\u89c6\u9891\u4f5c\u4e3a\u4e00\u4e2a\u53c2\u8003\u6807\u51c6\uff0c\u8fbe\u5230\u8fd9\u4e2a\u4ea4\u6d41\u7684\u6d41\u7545\u7a0b\u5ea6\u5c31\u975e\u5e38\u597d\u4e86\u3002Google \u6682\u65f6\u8fd8\u6ca1\u505a\u5230\u7684\u5b9e\u65f6\u6027\uff0c\u54ea\u5bb6\u516c\u53f8\u7387\u5148\u505a\u5230\u4e86\uff0c\u5c31\u662f\u7ade\u4e89\u529b\u3002<\/span><\/strong><\/p>\n<\/section>\nNano \u6a21\u578b\uff1a\u770b\u8d77\u6765\u4e0d\u592a\u7406\u60f3<\/span><\/strong><\/span><\/p>\nNano \u6a21\u578b\u503c\u5f97\u5173\u6ce8\uff0c1.8B \u548c 3.25B \u7684\u6a21\u578b\uff0c\u8fd8\u662f 4-bit \u91cf\u5316\u7684\uff0c\u5185\u5b58\u5360\u7528\u53ea\u6709\u5927\u7ea6 1 GB \u548c 2 GB\uff0c\u5728\u5927\u591a\u6570\u624b\u673a\u548c PC \u4e0a\u90fd\u53ef\u4ee5\u8dd1\u8d77\u6765\u3002\u4f46\u662f\u4ece\u8bc4\u6d4b\u62a5\u544a\u7684\u5f97\u5206\u4e0a\u770b\u8d77\u6765\u5e76\u4e0d\u662f\u7279\u522b\u7406\u60f3\uff0c\u4e0d\u77e5\u9053\u5b9e\u9645\u7528\u8d77\u6765\u6548\u679c\u5982\u4f55\u3002\u8fd9\u4e2a\u8bc4\u6d4b\u62a5\u544a\u4e0a\u9762\u5927\u91cf\u7684 SOTA \u5ba3\u79f0\u662f Google PaLI-X\uff0c\u5b83\u5728\u771f\u5b9e\u573a\u666f\u4e2d\u7684\u8868\u73b0\u5176\u5b9e\u662f\u4e0d\u5982 GPT-4V \u7684\uff0c\u8bc4\u6d4b\u62a5\u544a\u4e2d\u4e5f\u8bf4\u4e86 PaLI-X \u662f fine-tuned\uff0c\u56e0\u6b64 PaLI-X \u4e5f\u6709\u5237\u699c\u7684\u95ee\u9898\u3002<\/span><\/p>\n<\/p>\n
\u674e\u535a\u6770\uff1a\u4e2d\u79d1\u5927\u4e0e MSRA \u8054\u57f9\u8ba1\u7b97\u673a\u535a\u58eb\uff0cAI \u521b\u4e1a\u8005<\/span><\/p>\n\u539f\u6587\u94fe\u63a5\uff1a<\/span><\/p>\nhttps:\/\/www.zhihu.com\/question\/633684692\/answer\/3316416317<\/span><\/p>\n<\/h2>\n<\/h2>\n\u5f20\u4fca\u6797<\/strong><\/span><\/h2>\n\u6587\u672c\u53ef\u80fd\u662f\u5927\u6a21\u578b\u83b7\u53d6\u77e5\u8bc6\u7684\u4e3b\u8981\u6765\u6e90\u6e20\u9053<\/strong><\/span><\/h2>\n\u5982\u679c\u4ed4\u7ec6\u5206\u6790\u6280\u672f\u62a5\u544a\uff0c\u7ed3\u8bba\u5f88\u53ef\u80fd\u662f\u8fd9\u6837\u7684\uff1a\u5728\u6570\u5b66\u903b\u8f91\u7b49\u57fa\u7840\u5b66\u79d1\u80fd\u529b\u4e0a\u6765\u770b\uff0cGemini Ultra \u53ef\u80fd\u4e0d\u5982 GPT-4\uff0c\u591a\u6a21\u6001\u80fd\u529b\u4e0a\u5e94\u8be5\u5f3a\u4e8e GPT-4V<\/strong>\u3002<\/span><\/p>\n\u6280\u672f\u62a5\u544a\u8981\u70b9\u63d0\u70bc<\/span><\/strong><\/span><\/h3>\n\n\n- \n
\u6280\u672f\u62a5\u544a 60 \u9875\uff0c\u6ca1\u6709\u900f\u6f0f\u5177\u4f53\u6280\u672f\u7ec6\u8282\uff0c\u5927\u90e8\u5206\u662f\u8bc4\u6d4b\uff0c\u6280\u672f\u62a5\u544a\u4f5c\u8005\u5217\u8868\u5305\u542b 9 \u9875\u5185\u5bb9\uff0c\u8d85\u8fc7 700 \u4eba\uff0c\u5e94\u8be5\u63a5\u8fd1 OpenAI \u7684\u5458\u5de5\u603b\u6570\u4e86\u5427\u3002<\/span><\/p>\n<\/li>\n- \n
Gemini \u662f\u51e0\u79cd\u6a21\u6001\u4e00\u8d77\u8054\u5408\u4ece\u5934\u8bad\u7ec3\u7684\uff0c\u5305\u62ec\u6587\u672c\u3001\u56fe\u7247\u3001\u97f3\u9891\u3001\u89c6\u9891\u7b49\u3002\u8fd9\u4e0e\u76ee\u524d\u901a\u5e38\u7684\u591a\u6a21\u6001\u505a\u6cd5\u4e0d\u592a\u4e00\u6837\uff0c\u76ee\u524d\u7684\u591a\u6a21\u6001\u6a21\u578b\u4e00\u822c\u662f\u4f7f\u7528\u73b0\u6210\u7684\u8bed\u8a00\u5927\u6a21\u578b\u6216\u8005\u7ecf\u8fc7\u9884\u8bad\u7ec3\u8fc7\u7684\u56fe\u7247\u6a21\u578b\uff08\u6bd4\u5982 CLIP \u7684\u56fe\u7247\u7f16\u7801\u90e8\u5206\uff09\uff0c\u7136\u540e\u5229\u7528\u591a\u6a21\u6001\u8bad\u7ec3\u6570\u636e\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4e0a\u65b0\u7684\u7f51\u7edc\u5c42\u8bad\u7ec3\uff1b\u5982\u679c\u662f\u51e0\u4e2a\u6a21\u6001\u4ece\u5934\u5f00\u59cb\u4e00\u8d77\u8bad\u7ec3\uff0c\u90a3\u4e48\u6309\u7406\u8bf4\u5e94\u8be5\u90fd\u9075\u5faa next \u00a0token \u00a0prediction \u7684\u6a21\u5f0f\uff0c\u5c31\u5e94\u8be5\u662f LVM \u7684\u90a3\u4e2a\u8def\u5b50\uff0c\u5176\u5b83\u6a21\u6001\u7684\u6570\u636e\u6253\u6210 token\uff0c\u7136\u540e\u56fe\u7247\u3001\u89c6\u9891\u7b49\u5e73\u9762\u6570\u636e\u5148\u8f6c\u6362\u6210\u6bd4\u5982 16*16=256 \u4e2a token\uff0c\u7136\u540e\u641e\u6210\u4e00\u7ef4\u7ebf\u6027\u8f93\u5165\uff0c\u8ba9\u6a21\u578b\u9884\u6d4b next \u00a0token\uff0c\u8fd9\u6837\u5c31\u628a\u4e0d\u540c\u6a21\u6001\u5728\u8bad\u7ec3\u9636\u6bb5\u7edf\u4e00\u8d77\u6765\u3002<\/span><\/p>\n<\/li>\n- \n
\u6280\u672f\u62a5\u544a\u8bf4\u5e94\u8be5\u662f Decoder \u00a0only \u7684\u6a21\u578b\u7ed3\u6784\uff0c\u9488\u5bf9\u7ed3\u6784\u548c\u4f18\u5316\u76ee\u6807\u505a\u4e86\u4f18\u5316\uff0c\u4f18\u5316\u76ee\u7684\u662f\u5927\u89c4\u6a21\u8bad\u7ec3\u7684\u65f6\u5019\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u7684\u7a33\u5b9a\u6027\uff0c\u6240\u4ee5\u5927\u7ed3\u6784\u5e94\u8be5\u662f\u7c7b\u4f3c GPT \u7684 Decoder-only \u9884\u6d4b next \u00a0token prediction \u7684\u6a21\u5f0f\u3002\u76ee\u524d\u652f\u6301 32K \u4e0a\u4e0b\u6587\u3002<\/span><\/p>\n<\/li>\n- \n
Gemini \u00a0Nano \u5305\u542b\u4e24\u4e2a\u7248\u672c\uff1a1.8B \u9762\u5411\u4f4e\u7aef\u624b\u673a\uff0c3.25B \u9762\u5411\u9ad8\u7aef\u624b\u673a\u3002\u6587\u7ae0\u8bf4 Nano \u9996\u5148\u4ece\u5927\u6a21\u578b\u84b8\u998f\uff0c\u7136\u540e 4bit \u91cf\u5316\u3002\u6211\u8fd9\u91cc\u6709\u4e2a\u95ee\u9898\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u7528\u624b\u673a\u8c03\u7528 API \u7684\u65b9\u5f0f\u8c03\u7528\u670d\u52a1\u7aef\u7684\u6700\u5f3a\u6a21\u578b\u5462\uff1f\u80fd\u60f3\u5230\u7684\u4e00\u4e2a\u53ef\u80fd\u7684\u89e3\u91ca\u662f\u7528\u6237\u9690\u79c1\uff0c\u8fd9\u6837\u624b\u673a\u4e0d\u7528\u628a\u6570\u636e\u4f20\u5230\u4e91\u7aef\uff1b\u53e6\u5916\u4e00\u4e2a\u63a8\u7406\u6210\u672c\u4ece\u4e91\u7aef\u8f6c\u79fb\u5230\u4e86\u624b\u673a\uff0c\u80fd\u591f\u5927\u91cf\u8282\u7701\u63a8\u7406\u6210\u672c\u3002\u8fd8\u6709\u5176\u4ed6\u539f\u56e0\u4e48\uff1f\u6211\u4e00\u76f4\u4e0d\u592a\u7406\u89e3\u4e3a\u4f55\u8981\u505a\u585e\u5230\u624b\u673a\u91cc\u7684\u5927\u6a21\u578b\uff0c\u4e0d\u786e\u5b9a\u6838\u5fc3\u4f18\u70b9\u662f\u4ec0\u4e48\u3002<\/span><\/p>\n<\/li>\n- \n
\u4ece\u786c\u4ef6\u63cf\u8ff0\u90e8\u5206\u6765\u770b\uff0c\u610f\u601d\u662f\u52a8\u7528\u4e86\u524d\u6240\u672a\u6709\u7684 TPU \u96c6\u7fa4\uff0c\u6240\u4ee5\u63a8\u6d4b Gemini \u00a0Ultra \u7684\u6a21\u578b\u89c4\u6a21\u5e94\u8be5\u76f8\u5f53\u5927\uff0c\u731c\u6d4b\u5982\u679c\u662f MOE \u5927\u6982\u8981\u5bf9\u6807\u5230 GPT-4 1.8T \u7684\u6a21\u578b\u5bb9\u91cf\uff0c\u5982\u679c\u662f Dense \u6a21\u578b\u4f30\u8ba1\u8981\u5927\u4e8e 200B \u53c2\u6570\u3002\u8003\u8651\u5230\u5f15\u5165\u89c6\u9891\u97f3\u9891\uff08\u5f53\u7136\u662f\u6765\u81ea\u4e8e Youtube \u4e86\uff0c\u96be\u9053\u4f1a\u6765\u81ea TikTok \u4e48\uff09\u591a\u6a21\u6001\u6570\u636e\uff0c\u6240\u4ee5\u603b\u6570\u636e\u91cf*\u6a21\u578b\u53c2\u6570\uff0c\u4f1a\u662f\u975e\u5e38\u5de8\u5927\u7684\u7b97\u529b\u8981\u6c42\uff0c\u6280\u672f\u62a5\u544a\u8bf4\u53ef\u4ee5\u4e00\u5468\u6216\u8005\u4e24\u5468\u505a\u4e00\u6b21\u8bad\u7ec3\u3002<\/span><\/p>\n<\/li>\n- \n
\u8bad\u7ec3\u53ef\u80fd\u5206\u6210\u591a\u4e2a\u9636\u6bb5\uff0c\u6700\u540e\u9636\u6bb5\u63d0\u9ad8\u4e86\u9886\u57df\u6570\u636e\u7684\u6df7\u5408\u914d\u6bd4\uff0c\u731c\u6d4b\u5e94\u8be5\u6307\u7684\u662f\u903b\u8f91\u548c\u6570\u5b66\u7c7b\u7684\u8bad\u7ec3\u6570\u636e\u589e\u52a0\u4e86\u914d\u6bd4\uff0c\u76ee\u524d\u8c8c\u4f3c\u5f88\u591a\u8fd9\u4e48\u505a\u7684\uff0c\u5bf9\u4e8e\u63d0\u5347\u6a21\u578b\u903b\u8f91\u80fd\u529b\u6709\u76f4\u63a5\u5e2e\u52a9\u3002<\/span><\/p>\n<\/li>\n- \n
\u770b\u5b66\u79d1\u80fd\u529b\u6d4b\u8bd5\uff0c\u6280\u672f\u62a5\u544a\u6307\u6807\u6709\u4eba\u4e3a\u62d4\u9ad8\u7684\u503e\u5411\uff0c\u6bd4\u5982 MMLU\uff0c\u4ece 32 \u6b21\u6d4b\u8bd5\u91cc\u6295\u7968\u9009\u62e9\u6700\u4f18\u7b54\u6848\uff0c\u800c\u5bf9\u6bd4\u7684 GPT-4 \u5219\u4ec5\u4ece 5 \u4e2a\u6d4b\u8bd5\u91cc\u8fdb\u884c\u6295\u7968\uff0c\u8fd9\u4e2a\u5bf9\u6bd4\u660e\u663e\u4e0d\u516c\u5e73\u3002\u5f53\u4f8b\u5b50\u6570\u91cf\u90fd\u51cf\u5c11\u5230 5 \u4e2a\uff0cGemini \u00a0Ultra \u5f97\u5206 83.7%\uff0c\u4e0d\u5982 GPT-4 \u5f97\u5206 86.4%\uff0c\u9ad8\u4e8e GPT-3.5 \u7684 70%\u3002\u4ece\u6d4b\u8bd5\u5177\u4f53\u60c5\u51b5\u770b\uff0cgemini \u00a0ultra \u5e94\u8be5\u662f\u548c GPT-4 \u57fa\u672c\u6301\u5e73\u6216\u8005\u7a0d\u5fae\u5f31\u4e8e GPT-4 \u7684\uff0cGemini Pro \u548c Ultra \u5dee\u8ddd\u6bd4\u8f83\u5927\uff0c\u5e94\u8be5\u7565\u5fae\u5f3a\u4e8e GPT-3.5\uff1b\u800c\u4e14 Llama2 \u5728\u6570\u5b66\u3001\u63a8\u7406\u7b49\u65b9\u9762\u4e0e\u6700\u597d\u7684\u5927\u6a21\u578b\u6548\u679c\u5dee\u8ddd\u975e\u5e38\u660e\u663e\uff0c\u4e0d\u540c\u6d4b\u8bd5\u6307\u6807\u5dee\u8ddd 20 \u5230 40 \u5206\u4e4b\u95f4\uff1b<\/span><\/p>\n<\/li>\n- \n
\u4ece\u5b66\u79d1\u80fd\u529b\u6d4b\u8bd5\u6570\u636e\u770b\uff0c\u76ee\u524d\u5927\u6a21\u578b\u80fd\u529b\u5f88\u53ef\u80fd\u987a\u5e8f\u5982\u4e0b\uff1aGPT-4 \u7565\u5fae\u5f3a\u4e8e Geminni Ultra> Claude 2> inflection-2> GPT-3.5= Grok 1 >Llama2\u3002<\/span><\/p>\n<\/li>\n- \n
AlphaCode2 \u662f\u5728 Gemini Pro \u57fa\u7840\u4e0a\uff0c\u4f7f\u7528\u7f16\u7a0b\u7ade\u8d5b\u7684\u6570\u636e fine-tune \u51fa\u6765\u7684\uff0c\u6548\u679c\u63d0\u5347\u5f88\u660e\u663e\uff0c\u5728\u7f16\u7a0b\u7ade\u8d5b\u4e0a\u6392\u540d\u8d85\u8fc7 85% \u7684\u4eba\u7c7b\u9009\u624b\uff0c\u4e4b\u524d\u7684 AlphaCode1 \u8d85\u8fc7 50% \u7684\u4eba\u7c7b\u9009\u624b\uff1b<\/span><\/p>\n<\/li>\n- \n
Gemini Ultra \u5728\u591a\u6a21\u6001\u80fd\u529b\u65b9\u9762\uff0c\u5728\u51e0\u4e4e\u6240\u6709\u6d4b\u8bd5\u6570\u636e\u4e0a\u786e\u5b9e\u8981\u6bd4 GPT-4V \u5f3a\u4e00\u4e9b\u3002<\/span><\/p>\n<\/li>\n- \n
\u547d\u4ee4\u7406\u89e3\u65b9\u9762\uff1a\u548c GPT \u4e00\u6837\uff0c\u91c7\u7528\u591a\u6a21\u6001 instruct \u6570\u636e\u8fdb\u884c SFT+RM+RLHF \u4e09\u9636\u6bb5\uff0c\u8fd9\u91cc\u7684 RM \u90e8\u5206\u5728\u8bad\u7ec3\u6253\u5206\u6a21\u578b\u7684\u65f6\u5019\uff0c\u91c7\u7528\u4e86\u52a0\u6743\u7684\u591a\u76ee\u6807\u4f18\u5316\uff0c\u4e09\u4e2a\u76ee\u6807 helpfulness \u00a0factuality \u548c safety\uff0c\u731c\u6d4b\u5e94\u8be5\u662f\u5bf9\u4e8e\u67d0\u4e2a prompt\uff0c\u6a21\u578b\u751f\u6210\u7684\u7ed3\u679c\uff0c\u6309\u7167\u4e09\u4e2a\u6307\u6807\u5404\u81ea\u7ed9\u4e86\u4e00\u4e2a\u6392\u5e8f\u7ed3\u679c\u3002<\/span><\/p>\n<\/li>\n<\/ol>\n<\/section>\n\u4e00\u4e2a\u60b2\u89c2\u7684\u7ed3\u8bba<\/span><\/strong><\/span><\/h3>\n\u6700\u540e\u591a\u8bf4\u4e00\u53e5\uff0c\u4ece Gemini \u80fd\u591f\u63a8\u65ad\u51fa\u4e00\u4e2a\u60b2\u89c2\u7684\u7ed3\u8bba\u5982\u4e0b\uff1a<\/span><\/p>\n\u56e0\u4e3a\u5728 GPT-4V \u524d\u5927\u591a\u6570\u662f\u6587\u672c\u6a21\u578b\uff0c\u5f88\u591a\u4eba\u89c9\u5f97\u6587\u672c\u6a21\u578b\u7f3a\u4e4f Grounding\uff0c\u5c31\u662f\u6587\u672c\u62bd\u8c61\u8bed\u4e49\u548c\u771f\u5b9e\u7269\u7406\u5bf9\u8c61\u5bf9\u5e94\u4e0d\u8d77\u6765\uff0c\u5927\u6a21\u578b\u7406\u89e3\u4e0d\u4e86\u7269\u7406\u4e16\u754c\u7684\u77e5\u8bc6\uff0c\u800c\u89c6\u9891\u6570\u636e\u90a3\u4e48\u591a\uff0c\u5982\u679c\u5f15\u8fdb\u4e86\u540e\uff0c\u5927\u6a21\u578b\u4e0d\u4ec5\u80fd\u5efa\u7acb\u8d77 grounding\uff0c\u66f4\u91cd\u8981\u7684\u662f\u89c6\u9891\u6570\u636e\u8574\u542b\u4e86\u6bd4\u6587\u672c\u66f4\u591a\u7684\u77e5\u8bc6\uff0c\u6240\u4ee5\u5bf9\u5927\u6a21\u578b\u7684\u77e5\u8bc6\u50a8\u5907\u4f1a\u6709\u6781\u5927\u7684\u589e\u957f\u3002\u8fd9\u91cc\u53ef\u80fd\u5b58\u5728\u8bef\u89e3\u3002<\/span><\/p>\n\u4ece Gemini \u7684\u6548\u679c\u6765\u770b\uff0c\u4e8b\u5b9e\u53ef\u80fd\u5e76\u975e\u5982\u6b64\uff0cGemini \u591a\u6a21\u6001\u6548\u679c\u4e0d\u9519\uff0c\u5b83\u4e3b\u6253\u591a\u6a21\u6001\uff0c\u80af\u5b9a\u5f15\u5165\u4e86\u5c3d\u91cf\u591a\u7684\u89c6\u9891\u3001\u56fe\u7247\u4fe1\u606f\uff0c\u8fd9\u4e00\u65b9\u9762\u8bf4\u660e\u591a\u79cd\u6a21\u6001\u8054\u5408\u8bad\u7ec3\u786e\u5b9e\u6709\u7528\uff0c\u4f46\u662f\u5b83\u7684\u7528\u5904\u4e3b\u8981\u5728\u4e8e\uff1a\u628a\u6587\u672c\u62bd\u8c61\u6982\u5ff5\u548c\u7269\u7406\u5b9e\u4f53\u5f62\u8c61\u7684\u5bf9\u5e94 Grounding \u5efa\u7acb\u8d77\u6765\u4e86\uff0c\u4f46\u662f\u5728\u5927\u6a21\u578b\u7684\u4e16\u754c\u77e5\u8bc6\u548c\u5404\u79cd\u80fd\u529b\u50a8\u5907\u65b9\u9762\uff0c\u7ecf\u8fc7\u5927\u91cf\u89c6\u9891\u5f3a\u5316\u8fc7\u7684 Gemini \u751a\u81f3\u53ef\u80fd\u8fd8\u6bd4\u4e0d\u8fc7\u53ea\u7528\u6587\u672c\u8bad\u7ec3\u7684 GPT-4\u3002<\/span><\/p>\n\u8fd9\u4e00\u5207\u6307\u5411\u5982\u4e0b\u53ef\u80fd\uff1a\u5c31\u4e16\u754c\u77e5\u8bc6\u542b\u91cf\u6765\u8bf4\uff0c\u6587\u672c\u662f\u5927\u6a21\u578b\u83b7\u53d6\u77e5\u8bc6\u7684\u4e3b\u8981\u6765\u6e90\u6e20\u9053\uff0c\u89c6\u9891\u3001\u56fe\u7247\u6570\u636e\u5728\u8fd9\u65b9\u9762\u5bf9\u4e8e\u6587\u672c\u7684\u4e16\u754c\u77e5\u8bc6\u8865\u5145\u4f5c\u7528\u5fae\u4e4e\u5176\u5fae<\/strong>\uff0c\u89c6\u9891\u3001\u56fe\u7247\u548c\u6587\u672c\u591a\u6a21\u6001\u8bad\u7ec3\u7684\u4e3b\u8981\u4f5c\u7528\u662f\u5efa\u7acb\u8d77\u5b9e\u4f53\u6982\u5ff5\u53ca\u77e5\u8bc6\u62bd\u8c61\u8868\u8ff0\u548c\u5916\u5728\u7269\u7406\u5f62\u8c61\u7ed1\u5b9a\u5efa\u7acb grounding \u800c\u5df2\u3002\u9664\u6b64\u5916\uff0c\u65e0\u9700\u5bf9\u7c7b\u4f3c\u89c6\u9891\u7b49\u591a\u6a21\u6001\u6570\u5177\u6709\u66f4\u9ad8\u7684\u671f\u671b\u3002<\/span><\/p>\n\u672c\u8d28\u4e0a\uff0c\u76ee\u524d\u591a\u6a21\u6001\u5927\u6a21\u578b\u6548\u679c\u8fd8\u4e0d\u9519\uff0c\u662f\u5927\u6a21\u578b\u628a\u4ece\u6587\u672c\u4e2d\u5b66\u5230\u7684\u4e16\u754c\u77e5\u8bc6\u548c\u903b\u8f91\u80fd\u529b\uff0c\u7ecf\u8fc7 grounding \u7ed1\u5b9a\u5230\u5b9e\u4f53\u5916\u5728\u5f62\u8c61\u540e\uff0c\u5728\u591a\u6a21\u6001\u573a\u666f\u4e0b\u8bed\u8a00\u6a21\u578b\u628a\u4e30\u5bcc\u7684\u4e16\u754c\u77e5\u8bc6\u8fc1\u79fb\u7ed9\u4e86\u591a\u6a21\u6001\u6a21\u578b\uff0c\u662f\u6587\u672c\u6a21\u578b\u5e26\u7740\u591a\u6a21\u6001\u5728\u98de\uff0c\u800c\u4e0d\u662f\u53cd\u8fc7\u6765\u3002<\/span><\/p>\n\u5f20\u4fca\u6797\uff1a\u65b0\u6d6a\u5fae\u535a\u65b0\u6280\u672f\u7814\u53d1\u8d1f\u8d23\u4eba<\/span><\/p>\n\u539f\u6587\u94fe\u63a5\uff1a<\/span><\/p>\nhttps:\/\/www.zhihu.com\/question\/633684692\/answer\/3316675674<\/span><\/p>\n
\n\u5982\u679c\u4f60\u5173\u6ce8\u5927\u6a21\u578b\u9886\u57df\uff0c\u6b22\u8fce\u626b\u7801\u52a0\u5165\u6211\u4eec\u7684\u5927\u6a21\u578b\u4ea4\u6d41\u7fa4\uff0c\u6765\u4e00\u8d77\u63a2\u8ba8\u5927\u6a21\u578b\u65f6\u4ee3\u7684\u5171\u8bc6\u548c\u8ba4\u77e5\uff0c\u8ddf\u4e0a\u5927\u6a21\u578b\u65f6\u4ee3\u7684\u8fd9\u80a1\u6d6a\u6f6e\u3002<\/span><\/section>\n\n<\/section>\n<\/section>\n","protected":false},"excerpt":{"rendered":"\u4ee5\u4e0b\u6587\u7ae0\u6765\u6e90\u4e8e\u516c\u4f17\u53f7Founder Park\u00a0\uff0c\u4f5c\u8005Founder Park \u8c37\u6b4c\u5728 12 \u6708 6 \u65e5\u53d1\u5e03\u4e86 […]<\/p>\n","protected":false},"author":121,"featured_media":24508,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[82,81],"tags":[525,251,526],"class_list":["post-24498","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-82","category-81","tag-gemini","tag-251","tag-526"],"_links":{"self":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/24498","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/users\/121"}],"replies":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=24498"}],"version-history":[{"count":5,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/24498\/revisions"}],"predecessor-version":[{"id":24516,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/24498\/revisions\/24516"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/media\/24508"}],"wp:attachment":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=24498"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=24498"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=24498"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}