
{"id":6122,"date":"2025-08-14T22:37:29","date_gmt":"2025-08-14T14:37:29","guid":{"rendered":"https:\/\/infernews.com\/?p=6122"},"modified":"2025-08-15T03:38:35","modified_gmt":"2025-08-14T19:38:35","slug":"%e6%88%90%e7%82%ba%e9%a0%82%e5%b0%96-0-1-ai-%e7%a0%94%e7%a9%b6%e8%80%85%e7%9a%84-1-%e9%a0%85%e6%8a%80%e8%83%bd","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/%e6%88%90%e7%82%ba%e9%a0%82%e5%b0%96-0-1-ai-%e7%a0%94%e7%a9%b6%e8%80%85%e7%9a%84-1-%e9%a0%85%e6%8a%80%e8%83%bd\/","title":{"rendered":"\u6210\u70ba\u9802\u5c16 0.1% AI \u7814\u7a76\u8005\u7684 1 \u9805\u6280\u80fd"},"content":{"rendered":"\n<p><\/p>\n\n\n<figure class=\"wp-block-embed-youtube wp-block-embed is-type-video is-provider-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"lyte-wrapper\" title=\"1 SKILL To Become Top 0.1% AI Researcher\" style=\"width:853px;max-width:100%;margin:5px auto;\"><div class=\"lyMe\" id=\"WYL_dGCoIRGaKTw\" itemprop=\"video\" itemscope itemtype=\"https:\/\/schema.org\/VideoObject\"><div><meta itemprop=\"thumbnailUrl\" content=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdGCoIRGaKTw%2Fhqdefault.jpg\" \/><meta itemprop=\"embedURL\" content=\"https:\/\/www.youtube.com\/embed\/dGCoIRGaKTw\" \/><meta itemprop=\"duration\" content=\"PT26M25S\" \/><meta itemprop=\"uploadDate\" content=\"2025-08-12T12:37:06Z\" \/><\/div><div id=\"lyte_dGCoIRGaKTw\" data-src=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdGCoIRGaKTw%2Fhqdefault.jpg\" class=\"pL\"><div class=\"tC\"><div class=\"tT\" itemprop=\"name\">1 SKILL To Become Top 0.1% AI Researcher<\/div><\/div><div class=\"play\"><\/div><div class=\"ctrl\"><div class=\"Lctrl\"><\/div><div class=\"Rctrl\"><\/div><\/div><\/div><noscript><a href=\"https:\/\/youtu.be\/dGCoIRGaKTw\" rel=\"nofollow\"><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdGCoIRGaKTw%2F0.jpg\" alt=\"1 SKILL To Become Top 0.1% AI Researcher\" width=\"853\" height=\"460\" \/><br \/>Watch this video on YouTube<\/a><\/noscript><meta itemprop=\"description\" content=\"1 SKILL To Become Top 0.1% AI Researcher - EP.7 - Making LLMs Take Bytes As Input: Discover the underground technique elite AI researchers use to bypass tokenization limits\u2014feeding raw bytes directly into LLMs. This is how the top 0.1% are building faster, more efficient models. (AI prompt in video did not give good enough results) GitHub - https:\/\/github.com\/vukrosic\/bitstream-foundation-models-research Novita, cheapest 5090 and 4090 (choose spot billing for 50% off) - use my affiliate to help me get more compute for videos, you pay the same - https:\/\/novita.ai\/?ref=mjqyndm&amp;utm_source=affiliate Code LLM (GPT-5, DeepSeek, Qwen3, Llama4) from scratch - https:\/\/github.com\/vukrosic\/open-source-ai-2026 0:00 - The Goal: Byte-Based LLMs 1:31 - Planning the Research 2:27 - My &quot;Byte Latent Transformer&quot; Idea 4:50 - Understanding the Code 5:26 - How Text Becomes Bytes 7:02 - Grouping Bytes into Patches 9:12 - The Byte Encoder &amp; Embeddings 11:45 - The Mean Pooling Problem 14:48 - Generating Bytes (Decoding) 16:11 - Setting Up the GPU 16:53 - Designing the First Experiment 17:22 - What is Entropy-Based Patching? 23:24 - Experiment: Testing Entropy Thresholds 24:46 - Research Questions 26:08 - To Be Continued... - - Support my Free &amp; Open AI research - https:\/\/ko-fi.com\/vukai Help me do this full time\"><\/div><\/div><div class=\"lL\" style=\"max-width:100%;width:853px;margin:5px auto;\"><\/div><figcaption><\/figcaption><\/figure>\n\n\n<h3 class=\"wp-block-heading\">\u5f71\u7247\u8a73\u7d30\u4ecb\u7d39\uff1a1 SKILL To Become Top 0.1% AI Researcher &#8211; EP.7 &#8211; Making LLMs Take Bytes As Input<\/h3>\n\n\n\n<p>\u9019\u662f\u7cfb\u5217\u5f71\u7247\u7684\u7b2c 7 \u96c6\uff0c\u805a\u7126\u65bc AI \u7814\u7a76\u4e2d\u7684\u524d\u6cbf\u6280\u8853\uff1a\u8b93\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08LLMs\uff09\u76f4\u63a5\u4ee5\u4f4d\u5143\u7d44\uff08bytes\uff09\u4f5c\u70ba\u8f38\u5165\uff0c\u7e5e\u904e\u50b3\u7d71\u7684\u6a19\u8a18\u5316\uff08tokenization\uff09\u9650\u5236\u3002<\/p>\n\n\n\n<p>\u5f71\u7247\u7684\u76ee\u6a19\u662f\u5f15\u5c0e\u89c0\u773e\u4e00\u6b65\u6b65\u9032\u884c AI \u7814\u7a76\uff0c\u63a2\u8a0e\u5982\u4f55\u5728\u958b\u653e\u6e90\u78bc\u793e\u5340\u4e2d\u9818\u5148\u65bc\u50cf OpenAI \u6216 xAI \u9019\u6a23\u7684\u5de8\u982d\u3002Vuk Rosi\u0107 \u5f37\u8abf\uff0c\u9019\u9805\u6280\u8853\u80fd\u8b93\u6a21\u578b\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\uff0c\u56e0\u70ba\u6578\u4f4d\u4e16\u754c\u672c\u8cea\u4e0a\u5c31\u662f\u7531\u4f4d\u5143\u7d44\u7d44\u6210\u3002\u4ed6\u5206\u4eab\u81ea\u5df1\u7684\u7814\u7a76\u904e\u7a0b\uff0c\u5305\u62ec\u60f3\u6cd5\u69cb\u601d\u3001\u4ee3\u78bc\u89e3\u91cb\u3001\u5be6\u9a57\u8a2d\u8a08\uff0c\u4e26\u9f13\u52f5\u89c0\u773e fork GitHub \u5132\u5b58\u5eab\u53c3\u8207\u3002\u5f71\u7247\u4e2d\u63d0\u5230\uff0cElon Musk \u6700\u8fd1\u8868\u793a xAI \u5c07\u53d6\u4ee3 tokenizer\uff0c\u9019\u8b93\u9019\u9805\u7814\u7a76\u66f4\u5177\u6642\u6548\u6027\u3002<\/p>\n\n\n\n<p><!--more--><\/p>\n\n\n\n<p>\u4ee5\u4e0b\u662f\u5f71\u7247\u5167\u5bb9\u7684\u8a73\u7d30\u7d50\u69cb\u8207\u91cd\u9ede\u4ecb\u7d39\uff0c\u6211\u6703\u4f9d\u6642\u9593\u8ef8\u5206\u6bb5\u8aaa\u660e\uff0c\u8b93\u8b80\u8005\u66f4\u5bb9\u6613\u7406\u89e3\u3002\u5f71\u7247\u4ee5\u7b2c\u4e00\u4eba\u7a31\u6558\u8ff0\uff0cVuk Rosi\u0107 \u908a\u89e3\u8aaa\u908a\u5c55\u793a\u4ee3\u78bc\u3001\u63d0\u793a\u8a5e\uff08prompts\uff09\u548c AI \u5de5\u5177\u7684\u4f7f\u7528\uff08\u5982 Claude Opus 4.1\u3001Groq 4 \u7b49\uff09\uff0c\u4e26\u7a7f\u63d2\u4ed6\u7684\u500b\u4eba\u60f3\u6cd5\u548c\u7814\u7a76\u6311\u6230\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">0:00 &#8211; \u76ee\u6a19\uff1a\u57fa\u65bc\u4f4d\u5143\u7d44\u7684 LLMs<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6838\u5fc3\u6982\u5ff5<\/strong>\uff1a\u50b3\u7d71 LLMs \u4f7f\u7528 tokenizer \u5c07\u6587\u5b57\u8f49\u63db\u6210 token\uff0c\u9019\u6703\u9020\u6210\u6548\u7387\u554f\u984c\uff08\u5982\u8655\u7406\u591a\u8a9e\u8a00\u6216\u975e\u6587\u5b57\u8cc7\u6599\u6642\u7684\u9650\u5236\uff09\u3002Vuk \u63d0\u51fa\u76f4\u63a5\u8f38\u5165\u4f4d\u5143\u7d44\uff08raw bytes\uff09\uff0c\u8b93\u6a21\u578b\u5b78\u7fd2\u8655\u7406\u539f\u59cb\u8cc7\u6599\u3002\u9019\u80fd\u7bc0\u7701\u8a08\u7b97\u8cc7\u6e90\uff0c\u4e26\u9069\u7528\u65bc\u4efb\u4f55\u6578\u4f4d\u8cc7\u6599\uff08\u6587\u5b57\u3001\u5716\u50cf\u7b49\uff09\u3002<\/li>\n\n\n\n<li><strong>\u70ba\u4f55\u91cd\u8981<\/strong>\uff1a\u958b\u653e\u6e90\u78bc\u793e\u5340\u80fd\u9818\u5148\u5546\u696d\u516c\u53f8\uff0c\u63d0\u4f9b\u514d\u8cbb\u7814\u7a76\u6210\u679c\uff0c\u8b93\u5927\u5bb6\u53d7\u76ca\u3002\u4ed6\u63d0\u5230 tokenizer \u76ee\u524d\u6700\u4f73\u4f46\u6709\u7f3a\u9677\uff0c\u4eba\u5011\u6b63\u5c0b\u6c42\u66ff\u4ee3\u65b9\u6848\u3002<\/li>\n\n\n\n<li><strong>\u8cc7\u6e90\u5206\u4eab<\/strong>\uff1aGitHub \u5132\u5b58\u5eab\uff08https:\/\/github.com\/vukrosic\/bitstream-foundation-models-research\uff09\uff0c\u5305\u542b\u4ee3\u78bc\u548c\u7814\u7a76\u7b46\u8a18\u3002\u53e6\u5916\u63a8\u85a6 Novita AI \u7684 GPU \u670d\u52d9\uff08\u4f7f\u7528\u4ed6\u7684\u63a8\u85a6\u9023\u7d50\u53ef\u652f\u6301\u4ed6\u7372\u53d6\u66f4\u591a\u8a08\u7b97\u8cc7\u6e90\uff09\uff0c\u4ee5\u53ca\u53e6\u4e00\u500b\u958b\u653e\u6e90\u78bc AI \u5c08\u6848\uff08https:\/\/github.com\/vukrosic\/open-source-ai-2026\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">1:31 &#8211; \u7814\u7a76\u898f\u5283<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Vuk \u4f7f\u7528 AI \u5de5\u5177\uff08\u5982 ChatGPT\uff09\u4f5c\u70ba\u300c\u5c0e\u5e2b\u300d\u4f86\u898f\u5283\u7814\u7a76\u3002\u4ed6\u8a2d\u8a08\u4e00\u500b\u63d0\u793a\u8a5e\uff0c\u8b93 AI \u5e6b\u52a9\u4ed6\u601d\u8003\u7814\u7a76\u554f\u984c\u3001\u6a21\u677f\u548c\u6b65\u9a5f\u3002<\/li>\n\n\n\n<li><strong>\u91cd\u9ede<\/strong>\uff1a\u5f37\u8abf\u4eba\u985e\u512a\u52e2\u5728\u65bc\u76f4\u89ba\u601d\u8003\uff0c\u800c AI \u7528\u65bc\u7279\u5b9a\u4efb\u52d9\u3002\u7814\u7a76\u4e0d\u662f\u76f2\u76ee\u5beb\u4ee3\u78bc\uff0c\u800c\u662f\u5148\u6df1\u5165\u7406\u89e3\u554f\u984c\u3002<\/li>\n\n\n\n<li>\u4ed6\u4f7f\u7528 Notion \u9801\u9762\u7d44\u7e54\u7814\u7a76\uff0c\u6a19\u984c\u70ba\u300cReplacing Tokens with Bytes in LLMs Foundation Models\u300d\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2:27 &#8211; \u6211\u7684\u300cByte Latent Transformer\u300d\u60f3\u6cd5<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5275\u65b0\u9ede<\/strong>\uff1a\u9748\u611f\u4f86\u81ea Meta \u7684\u8ad6\u6587\uff0c\u63d0\u51fa\u300cByte Latent Transformer\u300d\uff1a\u4f7f\u7528\u5c0f\u578b\u795e\u7d93\u7db2\u8def\u9810\u6e2c\u4e0b\u4e00\u500b\u4f4d\u5143\u7d44\uff0c\u7576\u9810\u6e2c\u8b8a\u5f97\u56f0\u60d1\uff08perplexity \u9ad8\uff09\u6642\uff0c\u5c07\u4f4d\u5143\u7d44\u7fa4\u7d44\uff08patch\uff09\u7de8\u78bc\u6210\u6f5b\u5728\u5411\u91cf\uff08latent vector\uff09\uff0c\u985e\u4f3c token\u3002<\/li>\n\n\n\n<li><strong>\u512a\u52e2<\/strong>\uff1a\u5c0d\u7c21\u55ae\u4f4d\u5143\u7d44\u4f7f\u7528\u8f03\u5c11\u8a08\u7b97\uff0c\u5c0d\u8907\u96dc\u4f4d\u5143\u7d44\u4f7f\u7528\u66f4\u591a\u8a08\u7b97\u3002\u9019\u6a23\u80fd\u52d5\u614b\u5206\u914d\u8cc7\u6e90\uff0c\u63d0\u9ad8\u6548\u7387\u3002<\/li>\n\n\n\n<li>Vuk \u6bd4\u8f03\u591a\u500b AI \u6a21\u578b\u7684\u56de\u61c9\uff08\u5982 Claude Opus 4.1\u3001GPT-5 High\u3001Groq 4\uff09\uff0c\u8a8d\u70ba Opus \u89e3\u91cb\u6700\u7c21\u6f54\u6709\u7d50\u69cb\u3002\u4ed6\u5f37\u8abf\u8ddf\u96a8\u597d\u5947\u5fc3\u7814\u7a76\uff0c\u800c\u975e\u5b8c\u5168\u8907\u88fd Meta \u7684\u65b9\u6cd5\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">4:50 &#8211; \u7406\u89e3\u4ee3\u78bc<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6587\u5b57\u8f49\u4f4d\u5143\u7d44<\/strong>\uff1a\u793a\u7bc4\u5c07\u6587\u5b57\uff08\u5982 &#8220;hello&#8221;\uff09\u8f49\u6210 UTF-8 \u4f4d\u5143\u7d44\u5e8f\u5217\uff08e.g., [72, 101, 108, 108, 111]\uff09\u3002<\/li>\n\n\n\n<li><strong>\u8655\u7406\u591a\u4f4d\u5143\u7d44\u5b57\u7b26<\/strong>\uff1a\u6a21\u578b\u6703\u81ea\u52d5\u5b78\u7fd2\u8655\u7406 2-4 \u4f4d\u5143\u7d44\u7684\u5b57\u7b26\uff08\u5982\u975e ASCII \u7b26\u865f\uff09\uff0c\u7121\u9700\u984d\u5916\u8655\u7406\u3002<\/li>\n\n\n\n<li>Vuk \u6bd4\u8f03 AI \u6a21\u578b\u7684\u4ee3\u78bc\u89e3\u91cb\uff0c\u8b9a\u63da Opus 4.1 \u7684\u6e05\u6670\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">5:26 &#8211; \u5982\u4f55\u5c07\u6587\u5b57\u8f49\u6210\u4f4d\u5143\u7d44<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8a73\u7d30\u89e3\u91cb ASCII \u503c\uff080-255\uff09\u548c UTF-8 \u7de8\u78bc\u3002\u5f37\u8abf\u4f4d\u5143\u7d44\u8f38\u5165\u80fd\u8655\u7406\u4efb\u4f55\u8cc7\u6599\u985e\u578b\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">7:02 &#8211; \u5c07\u4f4d\u5143\u7d44\u7fa4\u7d44\u6210 Patch<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u56fa\u5b9a vs. \u52d5\u614b patch<\/strong>\uff1a\u907f\u514d\u56fa\u5b9a\u5927\u5c0f\uff08\u5982 8 \u4f4d\u5143\u7d44\uff09\uff0c\u6539\u7528\u57fa\u65bc perplexity \u7684\u52d5\u614b\u7fa4\u7d44\u3002\u7c21\u55ae\u5e8f\u5217\u7528\u5927 patch\uff0c\u8907\u96dc\u5e8f\u5217\u7528\u5c0f patch\u3002<\/li>\n\n\n\n<li><strong>\u586b\u5145\uff08Padding\uff09<\/strong>\uff1a\u4f7f\u7528 pad token \u586b\u5145\u4e0d\u6eff\u7684 patch\uff0c\u4e26\u8a0e\u8ad6\u662f\u5426\u5408\u4f75 end-of-sequence \u548c pad token \u4ee5\u7bc0\u7701\u7a7a\u9593\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">9:12 &#8211; Byte Encoder &amp; Embeddings<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7de8\u78bc\u5668<\/strong>\uff1a\u5c0f\u578b transformer\uff082 \u5c64\uff09\u5c07 patch \u8f49\u6210 embedding\u3002\u8a5e\u5f59\u5927\u5c0f\u70ba 258\uff08256 \u4f4d\u5143\u7d44 + 2 \u984d\u5916 token\uff09\u3002<\/li>\n\n\n\n<li><strong>\u81ea\u6ce8\u610f\uff08Self-Attention\uff09<\/strong>\uff1a\u6bcf\u500b\u4f4d\u5143\u7d44\u6ce8\u610f patch \u5167\u7684\u5176\u4ed6\u4f4d\u5143\u7d44\u3002<\/li>\n\n\n\n<li>Vuk \u8cea\u7591 embedding \u5927\u5c0f\uff08192 \u7dad\u5ea6\uff09\u662f\u5426\u5408\u9069\uff0c\u4f4d\u5143\u7d44\u6bd4 token \u66f4\u6a21\u7cca\uff0c\u53ef\u80fd\u9700\u8981\u66f4\u5927 embedding\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">11:45 &#8211; Mean Pooling \u7684\u554f\u984c<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Mean Pooling<\/strong>\uff1a\u5e73\u5747 patch \u5167\u6240\u6709 embedding \u6210\u55ae\u4e00\u5411\u91cf\uff0c\u4f46\u6703\u4e1f\u5931\u4f4d\u7f6e\u8cc7\u8a0a\u3002<\/li>\n\n\n\n<li><strong>\u6f5b\u5728\u6539\u9032<\/strong>\uff1a\u8a0e\u8ad6\u5176\u4ed6\u65b9\u6cd5\uff0c\u4f46\u5f37\u8abf\u7c21\u55ae\u6027\u512a\u5148\u3002\u6295\u5f71\u5c64\u5c07 192 \u7dad\u8f49\u6210\u6a21\u578b\u7684 384 \u7dad\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">14:48 &#8211; \u751f\u6210\u4f4d\u5143\u7d44\uff08Decoding\uff09<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u89e3\u78bc\u5668<\/strong>\uff1a\u5f9e patch embedding \u751f\u6210\u4f4d\u5143\u7d44\u5e8f\u5217\uff0c\u4f7f\u7528 teacher forcing\uff08\u8a13\u7df4\u6642\u63d0\u4f9b\u6b63\u78ba\u7b54\u6848\uff0c\u907f\u514d\u932f\u8aa4\u7d2f\u7a4d\uff09\u3002<\/li>\n\n\n\n<li><strong>\u81ea\u56de\u6b78\u751f\u6210<\/strong>\uff1a\u9010\u6b65\u9810\u6e2c\u4e0b\u4e00\u500b\u4f4d\u5143\u7d44\uff0c\u57fa\u65bc\u5148\u524d\u8f38\u51fa\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">16:11 &#8211; \u8a2d\u5b9a GPU<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u793a\u7bc4\u4f7f\u7528 Novita AI \u7684 RTX 5090\/4090 GPU\uff08spot \u5b9a\u50f9 50% \u6298\u6263\uff0c0.18 USD\/\u5c0f\u6642\uff09\u3002\u89e3\u91cb SSH \u9023\u7dda\u8a2d\u5b9a\uff0c\u9f13\u52f5\u4f7f\u7528\u4ed6\u7684\u63a8\u85a6\u9023\u7d50\u652f\u6301\u7814\u7a76\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">16:53 &#8211; \u8a2d\u8a08\u7b2c\u4e00\u500b\u5be6\u9a57<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528 Claude Opus \u5efa\u8b70\u5be6\u9a57\uff1a\u8abf\u6574 entropy \u95be\u503c\uff08threshold\uff09\u4f86\u6c7a\u5b9a patch \u5927\u5c0f\u3002<\/li>\n\n\n\n<li><strong>Entropy \u89e3\u91cb<\/strong>\uff1a\u6e2c\u91cf\u9810\u6e2c\u4e0d\u78ba\u5b9a\u6027\u3002\u4f4e entropy\uff08\u6613\u9810\u6e2c\uff09\u7e7c\u7e8c\u52a0\u5230 patch\uff0c\u9ad8 entropy \u958b\u59cb\u65b0 patch\u3002<\/li>\n\n\n\n<li>\u7bc4\u4f8b\uff1a\u5982 &#8220;hello world&#8221;\uff0centropy \u5728\u7a7a\u683c\u6216\u610f\u5916\u8a5e\u6642\u5347\u9ad8\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">17:22 &#8211; \u4ec0\u9ebc\u662f\u57fa\u65bc Entropy \u7684 Patching\uff1f<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8a73\u7d30\u516c\u5f0f\uff1a\u4f7f\u7528\u5c0f\u578b byte-level LM \u8a08\u7b97 entropy\uff08logits \u8f49\u6982\u7387\uff0c\u8a08\u7b97 -sum(p * log p)\uff09\u3002<\/li>\n\n\n\n<li><strong>\u5206\u88c2\u908f\u8f2f<\/strong>\uff1a\u9ad8 entropy \u6642\u5206\u88c2\uff0c\u6700\u5927 patch 16 \u4f4d\u5143\u7d44\u3002<\/li>\n\n\n\n<li>\u7bc4\u4f8b\uff1a\u91cd\u8907\u4f4d\u5143\u7d44\u4f4e entropy\uff0c\u96a8\u6a5f\u4f4d\u5143\u7d44\u9ad8 entropy\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">23:24 &#8211; \u5be6\u9a57\uff1a\u6e2c\u8a66 Entropy \u95be\u503c<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8a08\u5283\u8a13\u7df4 4 \u500b\u6a21\u578b\uff0c\u95be\u503c 1.5\u30012\u30012.5\u30013\u3002\u8ffd\u8e64\u640d\u5931\uff08loss\uff09\u3001\u68af\u5ea6\u7bc4\u6578\u3001\u8a18\u61b6\u9ad4\u4f7f\u7528\u7b49\u3002<\/li>\n\n\n\n<li><strong>\u7d71\u8a08<\/strong>\uff1a\u5e73\u5747 patch \u5927\u5c0f\u3001\u8b8a\u7570\u7b49\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">24:46 &#8211; \u7814\u7a76\u554f\u984c<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6700\u4f73\u7c92\u5ea6\uff1fpatch \u5927\u5c0f\u5982\u4f55\u5f71\u97ff\u5b78\u7fd2\u901f\u5ea6\uff1f\u54ea\u500b\u95be\u503c\u5e73\u8861\u8a9e\u7fa9\u4e00\u81f4\u6027\u548c\u7d30\u7c92\u5ea6\uff1f<\/li>\n\n\n\n<li>\u5047\u8a2d\uff1a\u4e2d\u7b49\u95be\u503c\uff082-2.5\uff09\u6700\u4f73\u3002<\/li>\n\n\n\n<li>Vuk \u63d0\u5230\u9650\u5236\uff08\u5982\u55ae GPU\uff09\uff0c\u4f46\u5f37\u8abf\u5f9e\u5c0f\u898f\u6a21\u958b\u59cb\u3002\u672a\u4f86\u9700\u5b78\u7fd2\u591a GPU \u5206\u6563\u8a13\u7df4\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">26:08 &#8211; \u5f85\u7e8c\u2026<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5f71\u7247\u56e0 Vuk \u75b2\u52de\u800c\u7d50\u675f\uff0c\u627f\u8afe\u7e8c\u96c6\uff08\u9023\u7d50\u5728\u63cf\u8ff0\u4e2d\uff09\u3002\u4ed6\u5f37\u8abf\u7406\u89e3\u55ae\u4e00\u65b9\u6cd5\u5c31\u80fd\u6d1e\u5bdf\u6574\u500b\u554f\u984c\uff0c\u9f13\u52f5\u89c0\u773e\u53c3\u8207\u7814\u7a76\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u6574\u9ad4\u8a55\u50f9\u8207\u555f\u767c<\/h4>\n\n\n\n<p>\u9019\u90e8\u5f71\u7247\u4e0d\u50c5\u662f\u6280\u8853\u6559\u7a0b\uff0c\u66f4\u662f AI \u7814\u7a76\u5165\u9580\u6307\u5357\uff0c\u9069\u5408\u5c0d LLMs \u611f\u8208\u8da3\u7684\u521d\u5b78\u8005\u548c\u7814\u7a76\u8005\u3002Vuk \u7684\u98a8\u683c\u89aa\u5207\u3001\u6b65\u9a5f\u660e\u78ba\uff0c\u5f37\u8abf\u958b\u653e\u6e90\u78bc\u548c\u597d\u5947\u5fc3\u9a45\u52d5\u7684\u7814\u7a76\u3002\u4ed6\u6279\u8a55 AI \u63d0\u793a\u8a5e\u6709\u6642\u4e0d\u7406\u60f3\uff08\u5982\u5f71\u7247\u6a19\u984c\u63d0\u5230\u7684 AI prompt \u7d50\u679c\u4e0d\u4f73\uff09\uff0c\u4f46\u5c55\u793a\u5982\u4f55\u8fed\u4ee3\u6539\u9032\u3002\u7f3a\u9ede\u662f\u7814\u7a76\u5c1a\u5728\u65e9\u671f\uff0c\u9700\u591a\u6708\u548c\u66f4\u591a\u4eba\u53c3\u8207\uff1b\u512a\u9ede\u662f\u63d0\u4f9b\u5b8c\u6574\u4ee3\u78bc\u548c\u5de5\u5177\uff0c\u8b93\u89c0\u773e\u80fd\u7acb\u5373\u5be6\u9a57\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u60a8\u60f3\u6df1\u5165\uff0c\u53ef\u76f4\u63a5 fork GitHub \u5132\u5b58\u5eab\uff0c\u6216\u89c0\u770b\u7cfb\u5217\u524d\u96c6\u5b78\u7fd2\u5f9e\u982d\u8a13\u7df4 LLMs\u3002\u5982\u679c\u6709\u7279\u5b9a\u90e8\u5206\uff08\u5982\u4ee3\u78bc\u6216 entropy \u8a08\u7b97\uff09<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5f71\u7247\u8a73\u7d30\u4ecb\u7d39\uff1a1 SKILL To Become Top 0.1% AI Researcher &#8211; EP.7 &#8211; Making LLMs Take Bytes As Input \u9019\u662f\u7cfb\u5217\u5f71\u7247\u7684\u7b2c 7 \u96c6\uff0c\u805a\u7126\u65bc AI \u7814\u7a76\u4e2d\u7684\u524d\u6cbf\u6280\u8853\uff1a\u8b93\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08LLMs\uff09\u76f4\u63a5\u4ee5\u4f4d\u5143\u7d44\uff08bytes\uff09\u4f5c\u70ba\u8f38\u5165\uff0c\u7e5e\u904e\u50b3\u7d71\u7684\u6a19\u8a18\u5316\uff08tokenization\uff09\u9650\u5236\u3002 \u5f71\u7247\u7684\u76ee\u6a19\u662f\u5f15\u5c0e\u89c0\u773e\u4e00\u6b65\u6b65\u9032\u884c AI \u7814\u7a76\uff0c\u63a2\u8a0e\u5982\u4f55\u5728\u958b\u653e\u6e90\u78bc\u793e\u5340\u4e2d\u9818\u5148\u65bc\u50cf OpenAI \u6216 xAI \u9019\u6a23\u7684\u5de8\u982d\u3002Vuk Rosi\u0107 \u5f37\u8abf\uff0c\u9019\u9805\u6280\u8853\u80fd\u8b93\u6a21\u578b\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\uff0c\u56e0\u70ba\u6578\u4f4d\u4e16\u754c\u672c\u8cea\u4e0a\u5c31\u662f\u7531\u4f4d\u5143\u7d44\u7d44\u6210\u3002\u4ed6\u5206\u4eab\u81ea\u5df1\u7684\u7814\u7a76\u904e\u7a0b\uff0c\u5305\u62ec\u60f3\u6cd5\u69cb\u601d\u3001\u4ee3\u78bc\u89e3\u91cb\u3001\u5be6\u9a57\u8a2d\u8a08\uff0c\u4e26\u9f13\u52f5\u89c0\u773e fork GitHub \u5132\u5b58\u5eab\u53c3\u8207\u3002\u5f71\u7247\u4e2d\u63d0\u5230\uff0cElon Musk \u6700\u8fd1\u8868\u793a xAI \u5c07\u53d6\u4ee3 tokenizer\uff0c\u9019\u8b93\u9019\u9805\u7814\u7a76\u66f4\u5177\u6642\u6548\u6027\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"googlesitekit_rrm_CAowvqSiDA:productID":"","footnotes":""},"categories":[27,146],"tags":[],"class_list":["post-6122","post","type-post","status-publish","format-standard","hentry","category-paper","category-146"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/6122","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=6122"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/6122\/revisions"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=6122"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=6122"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=6122"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}