Meta¤¬´û¸¤ÎÀ¸À®AI¤Ë¤¢¤ë¥È¡¼¥¯¥óÀ©¸Â¤ò¤Ï¤ë¤«¤Ë¾å²ó¤ë100Ëü¥È¡¼¥¯¥óĶ¤Î¥³¥ó¥Æ¥ó¥ÄÀ¸À®¤ò²Äǽ¤Ë¤¹¤ë¼¡À¤ÂåAI¥¢¡¼¥¥Æ¥¯¥Á¥ã¡ÖMegabyte¡×¤òȯɽ
Meta¤ÎAI¸¦µæ¥Á¡¼¥à¤¬¡¢100Ëü¥È¡¼¥¯¥ó¤òĶ¤¨¤ë¥Æ¥¥¹¥È¤ä²èÁü¡¢¥ª¡¼¥Ç¥£¥ª·Á¼°¤Ê¤É¤Î¥³¥ó¥Æ¥ó¥Ä¤òÀ¸À®¤Ç¤¤ë²ÄǽÀ¤ò»ý¤Ä¡ÖMegabyte(¥á¥¬¥Ð¥¤¥È)¡×¤Î¥¢¡¼¥¥Æ¥¯¥Á¥ã¤òÄ󾧤·¤Þ¤·¤¿¡£GPT-4¤äBard¤Ê¤É¤ÎAI¥â¥Ç¥ë¤Î´ðÈפȤʤäƤ¤¤ë¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Ï¡ÖTransformer¡×¤Ç¤¹¤¬¡¢Megabyte¤òAI¥â¥Ç¥ë¤ËºÎÍѤ¹¤ë¤³¤È¤Ç¡¢¤è¤ê¹âÅ٤ǽÏÎý¤·¤¿AI¥â¥Ç¥ë¤Î³«È¯¤¬²Äǽ¤Ë¤Ê¤ë¤ÈMeta¤Ï¼çÄ¥¤·¤Æ¤¤¤Þ¤¹¡£
https://doi.org/10.48550/arXiv.2305.07185
Meta AI Unleashes Megabyte, a Revolutionary Scalable Model Architecture - Artisana
https://www.artisana.ai/articles/meta-ai-unleashes-megabyte-a-revolutionary-scalable-model-architecture
OpenAI¤¬³«È¯¤ò¹Ô¤¦GPT-4¤Ê¤É¤Î¹âÀǽ¤ÊÀ¸À®AI¥â¥Ç¥ë¤Ï¡¢¥æ¡¼¥¶¡¼¤Ë¤è¤ëÊ£»¨¤ÊÆþÎϤòÍý²ò¤·¡¢Ä¹¤¤Ê¸¾Ï¤òÀ¸À®¤¹¤ë¤Î¤ËÌòΩ¤ÄTransformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤ò¥Ù¡¼¥¹¤È¤·¤Æ³«È¯¤¬¿Ê¤á¤é¤ì¤Æ¤¤¤Þ¤¹¡£¤·¤«¤·¡¢Meta¤ÎAI¸¦µæ¥Á¡¼¥à¤Ï¡ÖTransformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Ë¤Ï2¤Ä¤ÎÀ©¸Â¤¬¤¢¤ë¡×¤È»ØŦ¤·¤Æ¤¤¤Þ¤¹¡£AI¸¦µæ¥Á¡¼¥à¤Î»ØŦ¤¹¤ëÀ©¸Â¤Î¤Ò¤È¤Ä¤¬¡¢¥æ¡¼¥¶¡¼¤Ë¤è¤ëÆþÎϤÈAI¥â¥Ç¥ë¤Ë¤è¤ë½ÐÎϤ¬Ä¹¤¯¤Ê¤ë¤Ë¤Ä¤ì¤Æ¡¢Transformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Ç¤Ï·×»»Î̤¬Â¿¤¯¤Ê¤ê¡¢¡Ö¥È¡¼¥¯¥óÎ̤襤¥·¡¼¥±¥ó¥¹¤ò¸úΨŪ¤Ë½èÍý¤¹¤ë¤³¤È¤¬º¤Æñ¤Ë¤Ê¤Ã¤Æ¤¤¤¯¡×¤È¤¤¤¦ÅÀ¤Ç¤¹¡£
¤â¤¦¤Ò¤È¤Ä¤ÎÀ©¸Â¤Ï¡¢¸À¸ì¥â¥Ç¥ë¤¬°ìÏ¢¤Î¿ô³ØŪ±é»»¤ÈÊÑ´¹¤òÄ̤¸¤Æñ¸ì¤òÍý²ò¡¦½èÍý¤¹¤ë¤Î¤ò½õ¤±¤ë¡Ö¥Õ¥£¡¼¥É¥Õ¥©¥ï¡¼¥É¡¦¥Ë¥å¡¼¥é¥ë¥Í¥Ã¥È¥ï¡¼¥¯¡×¤¬¡¢Transformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Ç¤Ïʸ»ú¤Î¥°¥ë¡¼¥×¤´¤È¤ËÆÈΩ¤·¤ÆñÆÈÆ°ºî¤¹¤ë¤¿¤á¡¢·ë²Ì¤È¤·¤Æ·×»»Î̤¬Â¿¤¯¤Ê¤ê¡¢¥È¡¼¥¯¥óÎ̤襤¥·¡¼¥±¥ó¥¹¤Î½èÍý¤¬º¤Æñ¤Ë¤Ê¤Ã¤Æ¤·¤Þ¤¦¤È¤¤¤¦ÅÀ¤Ç¤¹¡£
¤³¤ì¤é¤ÎÍ×°ø¤«¤é¡¢¸½¾õ¤ÎTransformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Ï¡¢¥È¡¼¥¯¥óÎ̤襤Æþ½ÐÎϤò¸úΨŪ¤Ë½èÍý¤¹¤ë¤È¤¤¤¦ÅÀ¤Ç¸Â³¦¤Ë㤷¤Æ¤¤¤ë²ÄǽÀ¤¬»ØŦ¤µ¤ì¤Æ¤¤¤Þ¤¹¡£¤½¤³¤ÇMeta¤ÎAI¸¦µæ¥Á¡¼¥à¤Ï¡¢¤³¤ì¤é¤Î¸Â³¦¤ËÂн褷¡¢¥³¥ó¥Æ¥ó¥ÄÀ¸À®¤Î¿·¤¿¤Ê²ÄǽÀ¤ò°ú¤½Ð¤¹¤³¤È¤Ë¾ÇÅÀ¤òÅö¤Æ¤¿AI¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Î³«È¯¤ò¹Ô¤¤¤Þ¤·¤¿¡£
Meta¤¬³«È¯¤·¤¿Megabyte¤Ï¡¢Æþ½ÐÎϤ˴ØÏ¢¤¹¤ë¥·¡¼¥±¥ó¥¹¤ò¸Ä¡¹¤Î¥È¡¼¥¯¥ó¤Ç¤Ï¤Ê¤¯¡¢¡Ö¥Ñ¥Ã¥Á¡×¤´¤È¤Ëʬ³ä¤¹¤ëÆȼ«¤Î¥·¥¹¥Æ¥à¤òºÎÍѤ·¤Æ¤¤¤Þ¤¹¡£³Æ¥Ñ¥Ã¥Á¤Ï¥í¡¼¥«¥ë¤ÊAI¥â¥Ç¥ë¤Ë¤è¤Ã¤Æ½èÍý¤µ¤ì¡¢¤½¤Î¸å¥°¥í¡¼¥Ð¥ë¤ÊAI¥â¥Ç¥ë¤¬¤¹¤Ù¤Æ¤Î¥Ñ¥Ã¥Á¤òÅý¹ç¤·¤ÆºÇ½ªÅª¤Ê½ÐÎϤò¹Ô¤¤¤Þ¤¹¡£
Megabyte¤Î¥¢¥×¥í¡¼¥Á¤Ï¡¢¸½¾õ¤ÎAI¥â¥Ç¥ë¤¬Ä¾Ì̤·¤Æ¤¤¤ë²ÝÂê¤ËÂн褹¤ë¤â¤Î¤À¤È¤µ¤ì¤Æ¤ª¤ê¡¢Ã±°ì¤Î¥Õ¥£¡¼¥É¥Õ¥©¥ï¡¼¥É¡¦¥Ë¥å¡¼¥é¥ë¥Í¥Ã¥È¥ï¡¼¥¯¤¬Ê£¿ô¤Î¥È¡¼¥¯¥ó¤ò´Þ¤à¥Ñ¥Ã¥Á¤ÇʹԤ·¤ÆÆ°ºî¤¹¤ë¤³¤È¤Ç¡¢¤³¤ì¤Þ¤Ç¤Î²ÝÂê¤ò¹îÉþ¤Ç¤¤ë¤È¹Í¤¨¤é¤ì¤Æ¤¤¤Þ¤¹¡£
¥È¡¼¥¯¥ó¤Ç¤Ï¤Ê¤¯¥Ñ¥Ã¥Á¥Ù¡¼¥¹¤Ç¥·¥¹¥Æ¥à¤ò¹½ÃÛ¤¹¤ë¤³¤È¤Ç¡¢·×»»¤òÃ༡¹Ô¤¦½¾Íè¤ÎTransformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤È¤Ï°Û¤Ê¤ê¡¢Megabyte¤Ç¤Ï·×»»¤òÊÂÎ󤷤ƹԤ¦¤³¤È¤¬²Äǽ¤Ç¤¹¡£ÊÂÎó½èÍý¤ò¹Ô¤¦¤³¤È¤Ç¡¢Megabyte¤òÅëºÜ¤·¤¿AI¥â¥Ç¥ë¤Î¥Ñ¥é¥á¡¼¥¿¡¼¤¬Â¿¤¤¾ì¹ç¤Ç¤â¡¢Transformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¥Ù¡¼¥¹¤ÎAI¥â¥Ç¥ë¤è¤ê¤âÂçÉý¤Ê¸úΨ²½¤¬¼Â¸½¤Ç¤¤Þ¤¹¡£
¸¦µæ¥Á¡¼¥à¤¬¹Ô¤Ã¤¿¼Â¸³¤Ç¤Ï¡¢¥Ñ¥é¥á¡¼¥¿¡¼¿ô¤¬15²¯¤À¤Ã¤¿Megabyte¤Ï¡¢¥Ñ¥é¥á¡¼¥¿¿ô¤¬3²¯5000Ëü¤ÎTransformer¥â¥Ç¥ë¤è¤ê¤âÌó40¡ó®¤¯¥·¡¼¥±¥ó¥¹¤òÀ¸À®¤Ç¤¤ë¤³¤È¤¬¼¨¤µ¤ì¤Þ¤·¤¿¡£
¤µ¤é¤Ë¡¢GPT-4¤Ç¤Ï3Ëü2000¥È¡¼¥¯¥ó¡¢Anthropic¤Î¥Æ¥¥¹¥ÈÀ¸À®AI¡ÖClaude¡×¤Ç¤Ï10Ëü¥È¡¼¥¯¥ó¤¬À¸À®¤Î¾å¸Â¤À¤Ã¤¿¤Î¤ËÂФ·¡¢Megabyte¥â¥Ç¥ë¤Ç¤Ï¡¢120Ëü¥È¡¼¥¯¥ó¤òĶ¤¨¤ë¥·¡¼¥±¥ó¥¹¤ò½èÍý¤Ç¤¤ë¤³¤È¤¬ÌÀ¤é¤«¤Ë¤Ê¤ê¤Þ¤·¤¿¡£120Ëü¥È¡¼¥¯¥ó¤ò½èÍý²Äǽ¤ÊMegabyte¥â¥Ç¥ë¤Ï¡¢¥³¥ó¥Æ¥ó¥ÄÀ¸À®¤Î¿·¤¿¤Ê²ÄǽÀ¤ò³«¤¡¢¸½ºß¤ÎAI¥â¥Ç¥ë¤Î¸Â³¦¤òĶ¤¨¤ë¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Ë¤Ê¤ë¤³¤È¤¬´üÂÔ¤µ¤ì¤Æ¤¤¤Þ¤¹¡£
OpenAI¤Î¥ê¡¼¥ÉAI¥¨¥ó¥¸¥Ë¥¢¤Ç¤¢¤ë¥¢¥ó¥É¥ì¥¤¡¦¥«¡¼¥Ñ¥·¡¼»á¤ÏMegabyte¤ËÂФ·¤Æ¡ÖÂ絬ÌϸÀ¸ì¥â¥Ç¥ë¤Ë¤ª¤±¤ë¥È¡¼¥¯¥ó²½¤òÇѻߤǤ¤ë¤È¤¤¤¦ÅÀ¤Ç¡¢Megabyte¤ÏÍ˾¤Ç¤¹¡×¤È½Ò¤Ù¤Æ¤¤¤Þ¤¹¡£¤µ¤é¤Ë¡ÖChatGPT¤ÏÁϤŪ¤Ê¼¹É®¤äÍ×Ìó¤Ê¤É¤Î¥¿¥¹¥¯¤ËÍ¥¤ì¤Æ¤¤¤ë°ìÊý¡¢Í×Ì󤵤줿ʸ¾Ï¤òÉü¸µ¤¹¤ë¤Ê¤É¤Î¥¿¥¹¥¯¤ò¶ì¼ê¤È¤¹¤ë¤Î¤Ï¡¢¥È¡¼¥¯¥ó²½¤¬Í×°ø¤Ç¤¹¡×¤ÈÊó¹ð¤·¤Æ¤¤¤Þ¤¹¡£
Promising. Everyone should hope that we can throw away tokenization in LLMs. Doing so naively creates (byte-level) sequences that are too long, so the devil is in the details.
Tokenization means that LLMs are not actually fully end-to-end. There is a whole separate stage with¡Ä https://t.co/t240ZPxPm7— Andrej Karpathy (@karpathy) May 15, 2023
Meta¤ÎAI¸¦µæ¥Á¡¼¥à¤Ï¡¢Megabyte¥¢¡¼¥¥Æ¥¯¥Á¥ã¤¬²è´üŪ¤Êµ»½Ñ¤Ç¤¢¤ë¤³¤È¤òǧ¤á¤Ä¤Ä¤â¡¢ºÇŬ²½¤Î¤¿¤á¤Ë¤Ï¾¤Î¼êÃʤ¬¤¢¤ë²ÄǽÀ¤ò¼¨º¶¤·¤Æ¤¤¤Þ¤¹¡£¥Æ¥¯¥Î¥í¥¸¡¼·Ï¥Ë¥å¡¼¥¹¥á¥Ç¥£¥¢¤ÎArtisana¤Ï¡¢¡Ö¥Ñ¥Ã¥Áµ»½Ñ¤òºÎÍѤ·¤¿¤è¤ê¸úΨŪ¤Ê¥¨¥ó¥³¡¼¥À¥â¥Ç¥ë¤ä¡¢¥·¡¼¥±¥ó¥¹¤ò¤è¤ê¾®¤µ¤Ê¥Ö¥í¥Ã¥¯¤Ëʬ²ò¤¹¤ë¤¿¤á¤Î¥Ç¥³¡¼¥É¥â¥Ç¥ë¤Ê¤É¤ÎʬÌî¤Ë¤ª¤¤¤Æ¡¢½¾Íè¤ÎTransformer¥¢¡¼¥¥Æ¥¯¥Á¥ã¤Îµ¡Ç½¤ò³ÈÄ¥¤·¤Æ¡¢¼¡À¤Âå¤Î¥â¥Ç¥ë¤ËÂбþ¤Ç¤¤ë²ÄǽÀ¤¬¤¢¤ê¤Þ¤¹¡×¤È½Ò¤Ù¤Æ¤¤¤Þ¤¹¡£