{"id":4424,"date":"2024-03-21T17:33:45","date_gmt":"2024-03-21T09:33:45","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=4424"},"modified":"2024-04-20T14:25:17","modified_gmt":"2024-04-20T06:25:17","slug":"%e3%80%90%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%e3%80%91%e5%a4%a7%e6%a8%a1%e5%9e%8b%e4%b8%ad%e4%b8%80%e4%ba%9b%e6%9c%89%e6%84%8f%e6%80%9d%e7%9a%84idea","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/note-deep-learning\/4424\/","title":{"rendered":"\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011\u5927\u6a21\u578b\u4e2d\u4e00\u4e9b\u6709\u610f\u601d\u7684idea"},"content":{"rendered":"\n

\u8fd9\u4e2a\u7cfb\u5217\u5176\u5b9e\u4e5f\u548c\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011\u5927\u6a21\u578b\u4e2d\u7684\u4e00\u4e9bFAQ<\/a> \u4e00\u6837\uff0c\u6162\u6162\u66f4\u65b0\u3002<\/p>\n\n\n\n

\u6709\u4eba\u50ac\u66f4\u8bf7\u63091\uff0c\u6709\u4eba\u63d0\u95ee\u8bf7\u63092\uff0c\u6709\u4eba\u8ba8\u8bba\u8bf7\u6309\u5c71\uff0c\u6709\u4eba\u642c\u8fd0\u8bf7\u76f4\u63a5\u62e8110\uff01\uff01<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u6700\u65b0\u66f4\u65b0\u65f6\u95f4\uff1a2024.04.20<\/p>\n\n\n\n

\n\n\n\n

1.<\/strong>\u9884\u8bad\u7ec3\u65f6\uff0c\u5bf9\u65b0\u6570\u636e\u8fdb\u884c\u5fae\u8c03\u53ef\u80fd\u4f1a\u52a0\u5267\u5bf9\u9884\u8bad\u7ec3\u65f6\u6ca1\u6709\u5145\u5206\u4f9d\u636e\u7684\u65b0\u6570\u636e\u7684\u5e7b\u89c9\u3002<\/p>\n\n\n\n

\u4e5f\u5c31\u662f\u8bf4\uff0c\u6ca1\u5b66\u597d\u5c31\u662f\u6ca1\u5b66\u597d\uff0c\u540e\u671f\u662f\u6539\u4e0d\u8fc7\u6765\u7684\u3002<\/p>\n\n\n\n

\n

finetuning on newer data can aggravate hallucination for new data that is not well-grounded at pretraining time.<\/p>\n<\/blockquote>\n\n\n\n

from:http:\/\/arxiv.org\/abs\/2305.13169<\/a> https:\/\/www.youtube.com\/watch?v=hhiLw5Q_UFg<\/a><\/p>\n\n\n\n

2.<\/strong>\u6a21\u578b\u5b66\u7684\u662fA is B\uff0c\u4f46\u662f\u5f88\u96be\u63a8\u5e7f\u5230 B is A\u3002<\/p>\n\n\n\n

from:http:\/\/arxiv.org\/abs\/2309.12288<\/a><\/p>\n\n\n\n

3.<\/strong>transformer\u6a21\u578b\u6ca1\u6709\u201d\u521b\u9020\u201c\u7684\u529f\u80fd\uff0c\u5b83\u53ea\u80fd\u5f52\u7eb3\u6574\u7406\u5df2\u7ecf\u89c1\u8fc7\u7684\u6570\u636e\u3002\uff08\u4e0d\u8fc7\u662f\u7528\u6570\u5b66\u516c\u5f0f\u800c\u975e\u81ea\u7136\u8bed\u8a00\u505a\u7684\u5b9e\u9a8c\uff09\u3002<\/p>\n\n\n\n

\u6a21\u578b\u5176\u5b9e\u4e0d\u4f1a\u521b\u9020\u4e1c\u897f\uff0c\u53ea\u4f1a\u5bf9\u8bad\u7ec3\u6570\u636e\u8fdb\u884c\u603b\u7ed3\u62df\u5408\u3002Google\u7684\u4e00\u4e2a\u5b9e\u9a8c\uff0c\u53ef\u80fd\u98a0\u8986\u6240\u8c13\u7684AGI\uff1f\u4e0d\u8fc7\u5c40\u9650\u6027\u662f\u5b9e\u9a8c\u662f\u57fa\u4e8e\u6570\u5b66\u516c\u5f0f\u7684\uff0c\u5e76\u4e0d\u662f\u81ea\u7136\u8bed\u8a00\u3002<\/p>\n\n\n\n

from:http:\/\/arxiv.org\/abs\/2311.00871<\/a><\/p>\n\n\n\n

4.<\/strong>\u5927\u6a21\u578b\u662f\u901a\u8fc7\u6784\u5efa\u77e5\u8bc6\u56fe\u8c31\u6765\u8bb0\u5fc6\u77e5\u8bc6\u7684\u3002(\u563f\u563f \u7279\u610f\u548c2,3\u653e\u5728\u4e00\u8d77\uff09<\/p>\n\n\n\n

from:http:\/\/arxiv.org\/abs\/2309.14316<\/a><\/p>\n\n\n\n

\n\n\n\n

(2024.01.26\u66f4)<\/p>\n\n\n\n

5.<\/strong>\u9884\u8bad\u7ec3\u6a21\u578b\u5b9e\u9645\u4e0a\u5e76\u4e0d\u662f\u5728\u5b66\u4e60\u62df\u5408\u6570\u636e\u96c6\uff0c\u800c\u662f\u5728\u62df\u5408\u6570\u636e\u96c6\u7684\u751f\u6210\u8fc7\u7a0b\u3002<\/p>\n\n\n\n

\n

learning a CFG from the data\/learning addition rules from the data. Train on RNN(transformer) and find CFG algorithm<\/p>\n<\/blockquote>\n\n\n\n

intelligence = the generative process of human language within your brain<\/p>\n\n\n\n

from: Fu Yao<\/p>\n\n\n\n

6.<\/strong>\u6d8c\u73b0\u662f\u5426\u5b58\u5728\uff1f\u6d8c\u73b0\u662f\u5b87\u5b99\u4ece\u4e00\u76f4\u5b58\u5728\u7684\u73b0\u8c61\uff0c\u5316\u5b66\u53ef\u4ee5\u53cd\u8fc7\u6765\u88ab\u89c6\u4e3a\u7269\u7406\u5b9a\u5f8b\u7684\u7a81\u73b0\u5c5e\u6027\u3002\u751f\u7269\u5b66\uff08\u5305\u62ec\u751f\u7269\u8fdb\u5316\uff09\u53ef\u4ee5\u88ab\u89c6\u4e3a\u5316\u5b66\u5b9a\u5f8b\u7684\u7a81\u73b0\u5c5e\u6027\u3002\u540c\u6837\uff0c\u5fc3\u7406\u5b66\u53ef\u4ee5\u7406\u89e3\u4e3a\u795e\u7ecf\u751f\u7269\u5b66\u5b9a\u5f8b\u7684\u7a81\u73b0\u5c5e\u6027\u3002\u6700\u540e\uff0c\u4e00\u4e9b\u7ecf\u6d4e\u7406\u8bba\u5c06\u7ecf\u6d4e\u89c6\u4e3a\u5fc3\u7406\u5b66\u7684\u7a81\u73b0\u7279\u5f81\u3002<\/p>\n\n\n\n

\n

Chemistry can in turn be viewed as an emergent property of the laws of physics. Biology (including biological evolution) can be viewed as an emergent property of the laws of chemistry. Similarly, psychology could be understood as an emergent property of neurobiological laws. Finally, some economic theories understand economy as an emergent feature of psychology.<\/p>\n<\/blockquote>\n\n\n\n

\u4f46\u662f\u5927\u6982\u534a\u5e74\u524d\u6211\u4e5f\u786e\u5b9e\u770b\u8fc7\u4e00\u7bc7\u8bba\u6587\u8bf4\u6d8c\u73b0\u5176\u5b9e\u662f\u6d4b\u91cf\u5c3a\u5ea6\u4e0d\u7edf\u4e00\u5e26\u6765\u7684\uff0c\u6bd4\u59827b\u219213b\uff0c\u5dee\u4e86\u51e0b\uff0c13b\u219230b\u4e5f\u4e0d\u8fc7\u662f\u5341\u51e0b\uff0c\u800c\u518d\u5f80\u4e0a\u5c31\u5dee\u51fa\u4e86\u51e0\u5341b\u751a\u81f3\u767eb\u3002\u6a21\u578b\u5176\u5b9e\u5e76\u4e0d\u662f\u663e\u8457\u7684\u201c\u6d8c\u73b0\u201d\u80fd\u529b\uff0c\u800c\u662f\u5728\u8fd9\u4e2d\u95f4\u6301\u7eed\u7684\u6162\u6162\u53d8\u597d\uff0c\u4f46\u662f\u5e76\u6ca1\u6709\u83b7\u5f97\u771f\u5b9e\u7b54\u6848\u3002<\/p>\n\n\n\n

\u62ff\u6570\u5b66\u9898\u4e3e\u4f8b\uff1a\u8ba1\u7b971+1=\uff1f\uff0c7b\u53ef\u80fd\u8f93\u51fa100,13b\u53ef\u80fd\u662f50,30b\u53ef\u80fd\u662f5\uff0c\u6162\u6162\u7684\u76f4\u81f3100b\u8f93\u51fa\u4e862\u53ef\u4ee5\u88ab\u89c2\u6d4b\u5230\u6b63\u786e\uff0c\u90a3\u4e48\u8bf4\u6a21\u578b\u5728100b\u8fd9\u91cc\u53d1\u751f\u4e86\u6d8c\u73b0\u3002\u4f3c\u4e4e\u4e5f\u4e0d\u662f\u5f88\u5408\u7406\u3002(http:\/\/arxiv.org\/abs\/2304.15004<\/a>)\u3002\u4e0d\u8fc7\u4ece\u7ed3\u679c\u8bba\uff0c\u8fd9\u5c31\u662f\u6d8c\u73b0\u3002\u53ea\u8981\u6ca1\u6709\u4ece0~1\uff0c\u5373\u4f7f\u662f0.99\u5faa\u73af\uff0c\u4e5f\u662f0\u3002<\/p>\n\n\n\n

7.<\/strong>\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\u8fc7grounding\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<\/p>\n\n\n\n

\u4eceGemini\u7684\u699c\u5355\u8868\u73b0\u770b<\/p>\n\n\n\n

from: https:\/\/www.zhihu.com\/question\/633684692\/answer\/3316675674<\/a><\/p>\n\n\n\n

8.<\/strong>Transformer\u6a21\u578b\u53ef\u4ee5\u5f88\u7ec6\u81f4\u7684\u901a\u8fc7\u62df\u5408\u51fa\u4e09\u89d2\u6052\u7b49\u53d8\u6362\u8fd9\u7c7b\u6570\u5b66\u57fa\u7840\u516c\u5f0f\u6765\u5b8c\u6210\u4e0b\u6e38\u4efb\u52a1\u2014\u2014\u5373\u4f7f\u8fd9\u4e2a\u4efb\u52a1\u4e0d\u7528\u4e09\u89d2\u6052\u7b49\u89e3\u51b3\u3002\u603b\u4e4b\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7embedding\u8ba1\u7b97\u8f93\u5165\u6570\u503c\u7684cos\/sin\u503c\uff0c\u5c06\u8f93\u5165\u503c\u53d8\u6362\u5230\u6570\u5b66\u7684\u201c\u5706\u201d\u4e0a\uff0c\u968f\u540e\u901a\u8fc7\u4e09\u89d2\u6052\u7b49\u53d8\u6362\u7b49\u4e00\u7cfb\u5217\u64cd\u4f5c\uff0c\u6700\u540e\u8f93\u51fa\u8f93\u5165\u503c\u7684\u4e00\u4e9b\u5173\u7cfb\u3002\u5f88\u597d\u5bf9Transformer\u62df\u5408\u8fdb\u884c\u4e86\u53ef\u89e3\u91ca\u6027\u3002\u53cd\u6b63\u7ed9\u6211\u770b\u7684\u5f88\u8ff7\u60d1\uff0c\u9707\u60ca\u4e8eTransformer\u7adf\u7136\u53ef\u4ee5\u505a\u5230\u8fd9\u79cd\u7ea7\u522b\u7684\u62df\u5408\u3002<\/p>\n\n\n\n

from: http:\/\/arxiv.org\/abs\/2301.05217<\/a><\/p>\n\n\n\n

9.<\/strong>—delete—<\/p>\n\n\n\n

10.<\/strong>\u6a21\u578b\u7684warmup\u4f3c\u4e4e\u662f\u6ca1\u5fc5\u8981\u7684\uff1f\u4ece\u6211\u4eec\u548cbaichuan\u7684\u5b9e\u9a8c\u6765\u770b\u597d\u50cf\u786e\u5b9e\u662f\u8fd9\u6837\u7684\uff0c\u50cfLlama-MoE\u8fd9\u79cd\u5728\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u57fa\u7840\u4e0a\u65b0\u589e\u521d\u59cb\u5316gate\u7684\u505a\u6cd5\uff0c\u76f4\u89c9\u4e0a\u9700\u8981\u4f18\u5148\u8bad\u7ec3gate\uff0c\u8be2\u95ee\u4e86\u4e00\u4e0b\u4ed6\u4eec\uff0c\u4f3c\u4e4e\u51bb\u7ed3\/warmup\u4e5f\u662f\u6ca1\u5565\u533a\u522b\u7684\u3002<\/p>\n\n\n\n

from: https:\/\/arxiv.org\/abs\/2308.04014\u3001https:\/\/github.com\/pjlab-sys4nlp\/llama-moe\/issues\/55#issuecomment-1909369574<\/a><\/p>\n\n\n\n

11.<\/strong>Transformer\u6a21\u578b\u662f\u5728\u4e00\u4e2a\u70b9\u7684\u65f6\u5019\u7a81\u7136\u62e5\u6709\u201d\u667a\u80fd\u201c\u7684\u3002\u9996\u5148\u6b7b\u8bb0\u786c\u80cc\uff0c\u968f\u540e\u5728\u4e00\u4e2a\u70b9\u77ac\u95f4 \u201d\u5f00\u7a8d\u201c(grokking)\u5177\u6709\u6cdb\u5316\u6027\u3002<\/p>\n\n\n\n

from\uff1ahttps:\/\/pair.withgoogle.com\/explorables\/grokking\/<\/a> http:\/\/arxiv.org\/abs\/2301.05217<\/a><\/p>\n\n\n\n

\n\n\n\n

\uff082024..4.20\u66f4\uff09<\/p>\n\n\n\n

12. <\/strong>\u540c\u5c3a\u5bf8\u4e0b\uff0c\u4e00\u4e2a\u66f4\u6df1\u7684\u6a21\u578b\u4f1a\u6bd4\u66f4\u5bbd\u7684\u6a21\u578b\u6548\u679c\u66f4\u597d(\u8fd9\u4e5f\u5c31\u662f\u4e3a\u4ec0\u4e48post-norm\u8981\u6bd4pre-norm\u6548\u679c\u66f4\u597d\u7684\u539f\u56e0\uff0c\u4f46\u662fpre-norm\u8bad\u7ec3\u66f4\u52a0\u7a33\u5b9a)\u3002\u8fd9\u4e00\u70b9\u5e94\u8be5\u5728\u53e4\u8001\u7684DeepNorm\u8bba\u6587\u4e2d\u5c31\u63d0\u5230\u8fc7\u3002\u6700\u8fd1\u7684\u5f88\u591a\u7814\u7a76\u4e5f\u8bf4\u660e\u6df1\u5c42\u7684\u6a21\u578b\u5177\u5907\u66f4\u52a0\u4f18\u79c0\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n\n\n\n

from:https:\/\/arxiv.org\/pdf\/2312.17276.pdf<\/a><\/p>\n\n\n\n

13. <\/strong>\u5bf9\u4e8e\u4e00\u4e9b\u5c3a\u5bf8\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u901a\u8fc7\u5bf9\u7b2c\u4e00\u4e2aepoch\u8fdb\u884closs\u9009\u62e9(\u9009\u62e9loss\u8f83\u5927\u7684\u6570\u636e)\uff0c\u7ee7\u7eed\u8bad\u7ec3\u7b2c\u4e8c\u4e2aepoch\uff0c\u63a7\u5236\u6574\u4f53\u91c7\u6837\u7387\u5c0f\u4e8e50%\uff0c\u4f9d\u7136\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u6a21\u578b\u6027\u80fd\u3002\u4e0d\u8fc7\u8be5\u7ed3\u8bba\u662f\u8003\u8651\u5230\u5c0f\u5c3a\u5bf8\u6a21\u578b\u6574\u4f53\u5bb9\u91cf\u8f83\u5c0f\uff0c\u5bf9\u4e8e\u5927loss\u6570\u636e\u9057\u5fd8\u660e\u663e\uff0c\u4e0d\u786e\u5b9a\u80fd\u5426\u5411\u5927\u5c3a\u5bf8\u6570\u636e\u8fdb\u884c\u63a8\u5e7f\u3002\u4e0d\u8fc7\u6211\u4eec\u7684\u4e00\u4e9b\u5b9e\u9a8c\u4e5f\u8868\u660e\u591aepoch\u4f1a\u5bf9\u9a8c\u8bc1\u96c6loss\u4ea7\u751f\u79ef\u6781\u5f71\u54cd\u3002<\/p>\n\n\n\n

from:https:\/\/arxiv.org\/pdf\/2312.17276.pdf<\/a><\/p>\n\n\n\n

14.<\/strong> sft\u6ca1\u6709\u5b66\u4e60\u80fd\u529b\uff0c\u53ea\u662f\u5bf9\u9884\u8bad\u7ec3\u77e5\u8bc6\u7684\u6574\u7406\uff0c\u4ee5\u53ca\u683c\u5f0f\u5bf9\u9f50\u3002\u4e5f\u5c31\u662f\u8bf4\uff0csft\u57fa\u672c\u4e0d\u4f1a\u5b66\u4e60\u5230\u65b0\u77e5\u8bc6\u3002\u5982\u679csft\u7684\u4e16\u754c\u77e5\u8bc6\u548c\u9884\u8bad\u7ec3\u7684\u4e0d\u4e00\u81f4\uff0c\u4f1a\u5bfc\u81f4\u6a21\u578b\u5d29\u6e83\u3002<\/p>\n\n\n\n

from:http:\/\/arxiv.org\/abs\/2402.05119\u3001http:\/\/arxiv.org\/abs\/2402.18243<\/a><\/p>\n\n\n\n

15. <\/strong>scaling law\u76f8\u5173\u3002\u53ef\u4ee5\u901a\u8fc7\u4e00\u7cfb\u5217\u9884\u5b9e\u9a8c\uff0c\u63a8\u65ad\u51fa\u5927\u6a21\u578b\u5168\u6d41\u7a0b\u7684loss\u53d8\u5316\u3002\u540c\u65f6\u9a8c\u8bc1\u8fd9\u4e9bloss\u53d8\u5316\u662f\u8d85\u53c2\u65e0\u5173\u7684\uff08\u548c\u8bad\u7ec3\u6570\u636e\u3001tokenize\u548cmax len\u6709\u5173\u3002\u5728\u80fd\u591f\u83b7\u5f97\u8db3\u591f\u8bad\u7ec3<\/strong>\u7684\u60c5\u51b5\u4e0b\uff0c\u548clr\u3001lr\u8c03\u5ea6\u3001batch size\u5747\u65e0\u5173\uff09\uff0c\u540c\u65f6\u65b9\u6cd5\u5bf9\u9886\u57df\u5916\u7684\u6570\u636e\u96c6\u5177\u6709\u4e00\u5b9a\u7684\u666e\u9002\u6027\u3002 \u8fd9\u6709\u70b9\u7c7b\u4f3c\u4e8e23\u5e74\u4e0a\u534a\u5e74OpenAI\u8bf4\u4ed6\u4eec\u80fd\u591f\u9884\u6d4b\u6a21\u578b\u7684\u8bad\u7ec3\u66f2\u7ebf\u4e86\u3002<\/p>\n\n\n\n

from: http:\/\/arxiv.org\/abs\/2403.06563<\/a><\/p>\n\n\n\n

16.<\/strong> \u8bfe\u7a0b\u5b66\u4e60\/\u5206\u9636\u6bb5\u5b66\u4e60 \u4f3c\u4e4e\u662f\u4e00\u4e2a\u5341\u5206\u7b80\u5355\u5e76\u4e14\u6709\u6548\u7684\u5b66\u4e60\u65b9\u6848\u3002\u5f88\u591a\u8bba\u6587\u548c\u5b9e\u9a8c\u90fd\u6307\u660e\u5206\u9636\u6bb5\u8bad\u7ec3\uff08\u751a\u81f3\u4e0d\u9700\u8981\u4e25\u683c\u7684\u8bfe\u7a0b\u5b66\u4e60\uff09\u5bf9\u6a21\u578b\u6700\u7ec8\u6548\u679c\u662f\u6709\u589e\u76ca\u7684\u3002 \u4e0d\u8fc7\u6709\u8bf4\u6cd5\u5b66\u4e60\u7387\u7684\u591a\u6b21warmup\u548cdecay\u4f1a\u8ba9\u6a21\u578b\u66f4\u597d\u7684\u5b66\u4e60\u6570\u636e(http:\/\/arxiv.org\/abs\/2403.08763<\/a>)\uff0c\u4e0d\u592a\u786e\u5b9a\u662f\u5426\u662f\u56e0\u4e3a\u8fd9\u4e2a\uff0c\u56e0\u4e3awarmup\u5176\u5b9e\u4e5f\u4f1a\u52a0\u5267\u9057\u5fd8\u3002\u6709\u5173\u8fd9\u4e2a\u8bba\u6587\u7684\u66f4\u8be6\u7ec6\u4e00\u70b9\u7684\u8bf4\u660e\u770b 17<\/p>\n\n\n\n

from: http:\/\/arxiv.org\/abs\/2403.19270<\/a><\/p>\n\n\n\n

17. <\/strong>\u8fd8\u662f\u4e0a\u9762\u90a3\u4e2a\u8bba\u6587\uff0c\u4f46\u662f\u8fd9\u7bc7\u8bba\u6587\u91cc\u7684\u597d\u591a\u5b9e\u9a8c\u4e5f\u53ef\u4ee5\u8bf4\u660e\uff0c\u540c\u6837\u7684\u6570\u636e\u5206\u6279\u6b21\u8bad\u7ec3\u6548\u679c>\u653e\u5230\u4e00\u8d77\u8bad\u7ec3\uff0c\u751a\u81f3\u4f1a\u5b58\u5728\u8282\u7701\u6570\u767eB token\u5c31\u8fbe\u5230\u4e00\u6837\u7684\u4e0b\u6e38\u80fd\u529b\u7684\u60c5\u51b5\u3002<\/p>\n\n\n\n

\n

This is curious as these model have seen 150B and 100B fewer tokens of D1 than their respective baselines.\uff08D1 \u222a D0\uff09<\/p>\n<\/blockquote>\n\n\n\n

\u540c\u65f6\uff0c\u8fd8\u9a8c\u8bc1\u4e86\u591a\u9636\u6bb5\u91cc\u9002\u5f53\u7684replay\u6570\u636e\u53ef\u4ee5\u5728\u8f83\u5c11\u7684\u9057\u5fd8\u4ee3\u4ef7\u91cc\uff08\u5b8c\u7f8e\uff09\u62df\u5408\u65b0\u6570\u636e\uff0c\u76f8\u4f3c\u9886\u57df5%\uff0c\u504f\u5dee\u8f83\u592725%\u3002\u7b2c\u4e8c\u9636\u6bb5\u603btoken\u4e0d\u53d8\u7684\u60c5\u51b5\u4e0b\uff08\u6bd4\u5982\u7b2c\u4e8c\u9636\u6bb5\u8bad\u7ec3300B\uff0c\u6df7\u540820B\u524d\u4e00\u9636\u6bb5\u6570\u636e\u90a3\u5c31\u53ea\u4f7f\u7528280B\u65b0\u6570\u636e\uff09\uff0c\u53ef\u4ee5\u505a\u5230\u65e2\u51cf\u5c11\u9057\u5fd8\uff0c\u53c8\u4e0d\u5f71\u54cd\u65b0\u6570\u636e\u7684\u62df\u5408\uff08\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u4e0a\u9762\u8bf4\u8282\u7701\u767eB\u540c\u6027\u80fd\uff09\u3002<\/p>\n\n\n\n

from:http:\/\/arxiv.org\/abs\/2403.08763<\/a><\/p>\n\n\n\n

18. <\/strong>\u6700\u8fd1\u770b\u5230\u4e86ChatGPT\u503e\u5411\u4e8e\u9009\u62e942\uff0c\u7a81\u7136\u53d1\u73b0\u4e16\u754c\u6536\u675f\u4e86\u54c8\u54c8\u54c8\u54c8\u54c8\uff08\u867d\u7136ChatGPT\u9009\u62e942\u5927\u6982\u7387\u4e5f\u662f\u56e0\u4e3a\u8fd9\u4e2a\u5f69\u86cb\uff09\u3002<\/p>\n\n\n\n

\u4e3a\u5565\u5927\u5bb6\u7684\u968f\u673a\u79cd\u5b50\u603b\u559c\u6b22\u8bbe\u7f6e\u621042\u5462\uff1f\u8fd9\u6e90\u4e8e\u9053\u683c\u62c9\u65af\u00b7\u4e9a\u5f53\u65af\uff08Douglas Adams\uff09\u7684\u4f5c\u54c1\u300a\u94f6\u6cb3\u7cfb\u6f2b\u6e38\u6307\u5357\u300b\uff08The Hitchhiker’s Guide to the Galaxy\uff09\u3002\u5728\u8fd9\u4e2a\u7cfb\u5217\u4e2d\uff0c\u4e00\u4e2a\u8d85\u7ea7\u7535\u8111\u88ab\u8be2\u95ee\u201c\u751f\u547d\u3001\u5b87\u5b99\u4ee5\u53ca\u4e00\u5207\u201d\u7684\u7b54\u6848\uff0c\u5b83\u5f97\u51fa\u7684\u7b54\u6848\u662f\u201c42\u201d\u3002\u7136\u540e\u53ef\u80fd\u5c31\u4f5c\u4e3a\u4e00\u4e2a\u5c0f\u201c\u5f69\u86cb\u201d\u88ab\u4e00\u76f4\u6cbf\u7528\u3002<\/p>\n\n\n\n

\u7136\u540e\u968f\u774042\u51fa\u73b0\u7684\u6982\u7387\u53d8\u591a\uff0c\u6240\u4ee5\u8fd9\u79cd\u57fa\u4e8e\u7edf\u8ba1\u7684\u6a21\u578b\u5c31\u4f1a\u66f4\u503e\u5411\u4e8e\u9009\u62e942\uff1f<\/p>\n\n\n\n

from\uff1ahttps:\/\/twitter.com\/infobeautiful\/status\/1778059112250589561<\/a><\/p>\n\n\n\n

19.<\/strong> \u56de\u52306\u8bf4\u5230\u6d8c\u73b0\u53ef\u80fd\u662f\u56e0\u4e3a\u6d4b\u91cf\u6307\u6807\u975e\u8fde\u7eed\u9020\u6210\u7684\u3002\u6700\u8fd1\u667a\u8c31\u505a\u4e86\u4e00\u4e2a\u8fde\u7eed\u6027\u6307\u6807\u5b9e\u9a8c\u3002\u8bf4\u660e\u5373\u4f7f\u662f\u8fde\u7eed\u6027\u6307\u6807\u4f9d\u7136\u5b58\u5728\u6d8c\u73b0\u73b0\u8c61\u3002\u867d\u7136\u8fd9\u7bc7\u8bba\u6587\u4e3b\u8981\u65b9\u5411\u4e0d\u662f\u8fd9\u4e2a\u3002\u4f46\u662f\u6211\u8ba4\u4e3a\u8fd9\u662f\u8fd9\u7bc7\u8bba\u6587\u552f\u4e00\u62ff\u5f97\u51fa\u624b\u7684\u89c2\u70b9\u4e86emmm\u3002\u5176\u4f59\u7684\u4ec0\u4e48loss\u624d\u662f\u6d8c\u73b0\u7684\u5173\u952e\uff0c\u8bf4\u6cd5\u6709\u70b9\u7275\u5f3a\uff0c\u6574\u4f53\u7ed3\u679c\u4f9d\u7136\u8fd8\u662f\u5448\u73b0\u6a21\u578b\u89c4\u6a21\uff08\u8d85\uff09\u5f3a\u76f8\u5173\u7684\u8d8b\u52bf\u3002<\/p>\n\n\n\n

from:https:\/\/arxiv.org\/abs\/2403.15796<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

\u8fd9\u4e2a\u7cfb\u5217\u5176\u5b9e\u4e5f\u548c\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011\u5927\u6a21\u578b\u4e2d\u7684\u4e00\u4e9bFAQ \u4e00\u6837\uff0c\u6162\u6162\u66f4\u65b0\u3002 \u6709\u4eba\u50ac\u66f4\u8bf7\u63091\uff0c\u6709\u4eba\u63d0\u95ee\u8bf7\u6309 […]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":true,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_mi_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[9],"tags":[],"views":6284,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/4424"}],"collection":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/comments?post=4424"}],"version-history":[{"count":9,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/4424\/revisions"}],"predecessor-version":[{"id":4538,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/4424\/revisions\/4538"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=4424"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=4424"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=4424"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}