ChatGPTãæãããããããããªãã¨ã人éããè³¢ããã£ã¦ããã®ãè¦ã¦ã·ã³ã®ã¥ã©ãªãã£ã¨ããè¨èã使ã人ãå¢ããããã«æãã¾ãããéã«ãã·ã³ã®ã¥ã©ãªãã£ã¯æ¥ãªãã®ã§ã¯ã¨ããæããå¼·ããã¦ãã¾ãã
ã¾ãããã®æç« ã§ã®ã·ã³ã®ã¥ã©ãªãã£ããªã«ãã¨ãã話ã§ãããã¬ã¤ã»ã«ã¼ãã¯ã¤ã«ããã·ã³ã®ã¥ã©ãªãã£ã¯è¿ããã®1ç« ã®çµããã§ãããããããç¹ç°ç¹ã ãã¨ãã£ã¦ããç¹ç°ç¹ããã®ã·ã³ã®ã¥ã©ãªãã£ã§ãã
ãã®ç¹ç°ç¹ã¯åã«AIã人éããè³¢ããªãã¨ããã ãã§ã¯ããã¾ããã人éããè³¢ããªãã ãã ã¨ã便å©ãªéå ·ãå¢ããã ããªã®ã§ã大é¨ãããã»ã©ã®å¤åã¯èµ·ãã¾ããã人ã®ä»äºã奪ãã¨ãã£ã¦ããè¸æ°æ©é¢ã»ã©ã§ããªãã§ãããå°å·æ©ãªã©ã¨ä¸¦ãã§ã人ã®çæ´»ãå¤ãã転æç¹ã«ãããããã ã²ã¨ã¤ã®ç¹ãããããã·ã³ã®ã¥ã©ãªãã£ã«ã¯ ãªãã¾ããã(çæ´»ã«å¤åã¯ããã¨æãã¾ã)
ã«ã¼ãã¯ã¤ã«ãã·ã³ã®ã¥ã©ãªãã£ã¨è¨ã£ã¦ããã®ã¯ã人工ç¥è½ãèªå¾çã«ãããã人工ç¥è½ãä½ããããã«ãªãã¨äººéã®æã«ããé²åãããã¯ããã«éãçºå±ãè¡ããããã¨ã«ãªããææ°é¢æ°ã®ææ°é¢æ°ã®ãããªççºçãªé²åã訪ããã¨ãããã¨ããã§ããããã人éã®æã«ããé²åã¨ã¯éããã®ã«ãªãã®ã§ããã®è»¢æç¹ãã·ã³ã®ã¥ã©ãªãã£ã¨å¼ãã§ãã¾ãã
ãã ã人工ç¥è½ã®å½¢ãChatGPTã®ãããªãã®ã«ãªã£ãã®ã§ããã·ã³ã®ã¥ã©ãªãã£ã¯è¿ããã®1ç« ã«æ¸ããã¦ããåæãå´©ãã¦ãã¾ãã
ã¾ã次ã®ç¹ãããã¯ã«ã¼ãã¯ã¤ã«ãæ³å®ãã人工ç¥è½ããã½ã¼ã¹ã³ã¼ãã§è¨è¿°ããããã¸ãã¯ã«ãããã®ã§ãããã¨ã表ãã¦ãã¾ãã
æ©æ¢°ãã人éã®ãã¤è¨è¨æè¡è½åãç²å¾ããã°ãé度ã容éã¯äººéã®ãããã¯ããã«è¶ ããæ©æ¢°èªèº«ã®è¨è¨ï¼ã½ã¼ã¹ã³ã¼ãï¼ã«ã¢ã¯ã»ã¹ããèªèº«ãæä½ããè½åããã¤ãã¨ã«ãªãã
ããã©ãChatGPTãªã©ãã¾ã®äººå·¥ç¥è½ã®ãã¢ã¯æå ¥ãããã¼ã¿ã§ããã½ã¼ã¹ã³ã¼ãèªä½ã¯ãããããGPT1ã®ããããã»ã¨ãã©å¤ãã£ã¦ãªãã¨æããã¾ãã ã¾ããã½ã¼ã¹ã³ã¼ãã§ã¯ãªããã¥ã¼ã©ã«ãããã®æ§é ã ã¨ãã¦ãLLMã®å¦ç¿ã®ããã«ã¯ãã®äººå·¥ç¥è½ãåãã¦ããã³ã³ãã¥ã¼ã¿ã®ä½åãã®ã³ã³ãã¥ã¼ã¿ãå¿ è¦ã«ãªããããç°¡åã§ã¯ããã¾ããã
ããã¦æ¬¡ã®ç¹ã
æ©æ¢°ã¯ãããèªèº«ã®è¨è¨ãçµã¿æ¿ãã¦ãæ§è½ãééãªãå¢å ããããã¨ãã§ããã
LLMã®æ§è½ããã¸ãã¯ã®æ¹åã§ã¯ãªãã¢ãã«è¦æ¨¡ã¨å¦ç¿ãã¼ã¿ãè¨ç®ãªã½ã¼ã¹ã®å¤§ããã§æ±ºã¾ã以ä¸ã¯ãæ§è½ãééãªãå¢å ã¨ããã®ã¯é£ãããã§ãã
æ¨è«æãã¤ã¾ãå®éã«ãããã¹ããã段éã§ã¯ã¢ãã«ã®è»½éåãRWKVã®ãããªä»çµã¿ãããã¾ãããå¦ç¿æã§ã¯ããã¾ã§è»½éåã§ãã¦ãªãããã«æãã¾ãã(RWKVãå¦ç¿æã¯Transformerçã¨ãã話)
ããã¦ãä¾ãã°GPT4ã§ããã°ãããã1TBãããã®ã¡ã¢ãªã使ããGPUã使ãã¨æãã®ã§ãããããã2TBã®ã¡ã¢ãªãè¼ã£ãGPUãå¿
è¦ã¨ãªã£ããã人工ç¥è½åä½ã§ã¯ã©ãã«ããªãã¾ãããã
ã¾ããå¦ç¿ã«ã¯ãã¼ã¿ã»ã³ã¿ã¼ä¸è»ãããã®ã³ã³ãã¥ã¼ã¿ãå¿
è¦ã§ããã·ã³ã®ã¥ã©ãªãã£ã¨ãªã人工ç¥è½ã¯ãã®æç¹ã§æé«ã®ãã¼ã¿ã»ã³ã¿ã¼ã§å¦ç¿ããã¦ããã¨æãã®ã§ããããã«ãã¼ã¿ã»ã³ã¿ã¼ã建ã¦ãå¿
è¦ãããã¾ããå°ãªãã¨ããæ°ããGPUãè¼ã£ãè¨ç®æ©ã¯ã©ã¹ã¿ãä½ãå¿
è¦ãããã¾ãã
ãããããã®è²»ç¨ã©ãããã®ãã¨ãããã¨ãèããã¨ãééãªãã¨ããã®ã¯ã¡ãã£ã¨ç¡çã§ããã
人工ç¥è½ãèªåèªèº«ãå¼·åããããã®ãµã¼ãã¼ä»£ã稼ãããã«TikTokãYouTubeã®åç»ãä½ã£ã¦ç¨¼ãã§ãã®ãæ³åããã¨ãã¡ãã£ã¨ç¬ã£ã¦ãã¾ããã©ã
ããã¦ãããç¨åº¦ã®è¦æ¨¡ã«ãªãã¨ãé»åè¨è¨çã«éçããã¦ãã¾ãã¾ãã
â» è¿½è¨ æ ¸èåã®è©±ãã§ã¦ã¾ãããçºé»éã®åé¡ã§ã¯ãªã建ç©ã«å¼ãè¾¼ãã§æ°ãæé«è² è·ãããã¤ã¤24æéè½ã¡ãªããã管çã§ããé»åéã®åé¡ã§ããçºé»éã§ããã°ç«åçºé»1åºã§ã¾ããªããç¨åº¦ã§ãã(ã¨ãã話ãå¾è¿°ã®æ¾å²¡å
çãããã¦ãã¾ããããããç±ç®¡çã®åé¡ã§ã¯ãªããã¨ã)
æå¾ã«æ¬¡ã®æ®µããããã·ã³ã®ã¥ã©ãªãã£ãã·ã³ã®ã¥ã©ãªãã£ã§ããæéè¦é ç®ã§ãã
æ©æ¢°ã®ç¥è½ãèªèº«ã®è¨è¨ãç¹°ãè¿ãæ¹åãããµã¤ã¯ã«ã¯ãã©ãã©ãéããªãã
ãã©ãããã¾ã§è¦ãã¨ããªããªãã©ãã©ãéãã¨ããã®ã¯é£ããã§ããããããã人éãå¹çããåããç¨åº¦ã§ã¯ãªãã§ããããã
ããè³¢ã人工ç¥è½ã¨ããã®ã¯ããã大éã®å¦ç¿ãã¼ã¿ããã大éã®è¨èªã¢ãã«ã«æå
¥ãã¦ãããã¨ã«ãªãã®ã§ãæ°ãæã®å¦ç¿æéããããã¾ããããã¦ãæ°ãæã®å¦ç¿ãã¾ããããã«è³éã®ç¢ºä¿ãé»åã®ç¢ºä¿ãªã©ç¤¾ä¼ã¨ã®ãããããå¿
è¦ã«ãªãã¾ãã
ããã·ã³ã®ã¥ã©ãªãã£AIãéåã³ã³ãã¥ã¼ã¿ã§åãã¦ããã°ãããè³¢ãAIãä½ãããã«éåããããå¢ããå¿ è¦ãããã¨ãªãã°ç°¡åãªãã¨ã§ã¯ããã¾ããã
ã¨ãããã¨ã§ã人工ç¥è½ã®çºå±ã¯äººå·¥ç¥è½ã§éãã¦è¡ããããã§ã¯ãªã人é社ä¼ã¨ã®é¢ãããå¿
è¦ã«ãªãã®ã§ãã·ã³ã®ã¥ã©ãªãã£ã¨è¨ããã»ã©ã®é²åã¹ãã¼ãã®å¤åã¯èµ·ããªããã¨ããçµè«ã§ãã
ã¾ããå¯è½æ§ã¯ã¼ãã§ã¯ãªããã§ããã©ãã
ã¨ããã§ãGPT4ã®ãã¨ã¨ããã®ãæ ¼æ®µã«æ§è½åä¸ããã®ã¯ãããªã«ç°¡åãããªãã®ã§ã¯ãã¨ãã話ãããããåºã¦ãã¦ãã¾ãã
ãã®ã¨ã³ããªã®ã¨ãã«ã¯ãå¦ç¿ãã¼ã¿ã®æ¯æ¸ã主ã«æãã¾ããã
大規模言語モデルはこれ以上賢くならず庶民的になっていく - きしだのHatena
è¨ç®è³æºã®è¦³ç¹ããã¯æ¾å²¡å
çããããã話ãããã¦ãã¾ãã
ã¨ããã§BERTããGPT4ã¾ã§æ¥éãªé²åãéããLLMã®é²åã§ãããä»å¾ã¯å°ãªãã¨ãäºåå¦ç¿ã¢ãã«ã®é²åã®ã¹ãã¼ãã¯å¨å²ã®æå¾ ã«åãã¦å ¨è¬çã«ã¯é ããªãã§ããããããã¯ãå¦ç¿ã«ç¨ããã¹ãã³ã³ãä¸çã®æ±ç¨ã¹ãã³ã³ã®ãããã¬ãã«ã«æ¥éã«è¿½ãã¤ãã¦ãã¾ã£ãããã§ãã
— Satoshi Matsuoka (@ProfMatsuoka) 2023å¹´4æ11æ¥
ããã¦OpenAIã®ãµã ã»ã¢ã«ããã³èªèº«ãããã¼ã¿ã»ã³ã¿ã¼æ§ç¯ã®ç©ççãªå¶ç´ãæ§ç¯é度ãããã¦è¨èªã¢ãã«ã大ããããã¨ãã®æ§è½åä¸ãæ¸å°ãã¦ãããã¨ãªã©ãããã¦ãGPT4以éã¯å¥ã®æ¹æ³ãã¨ãå¿
è¦ãããã¨ãã話ããã¦ãã¾ãã
https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/
ãã¯ããä¸ã«æããã¨ã³ããªã§æ¸ããããã«ãã²ã¨ã¤ã®ã¨ã¦ãè³¢ãã¢ãã«ã§å ¨é¨ã®éè¦ãè³ãå¿ è¦ã¯ãªãã¦ãã»ã©ã»ã©ã«è³¢ãè¨èªã¢ãã«ãããã° ãã¨ã¯ãã¸ãã¯ãã¢ããªã±ã¼ã·ã§ã³ã®ä½ããã¿ã追å å¦ç¿ãªã©ã§ã ãããã®éè¦ã¯è³ããã®ã§ããã£ã¨è³¢ãã¢ãã«ãä½ããã¨ããæ¹åã§ã¯ãªãã使ããããã¨ããæå³ã§ã庶æ°çãªæ¹åã§ã¢ãã«ãä½ããã¦ããããããªãããã¨æãã¾ãã
ãã®ã¨ã³ããªã§ãChatGPTã«ããã®çºè¨ã¯æ£ããã§ããï¼ãã¨å°ããã¨èª¤ããè¦ã¤ãããã¨ãå¤ãã®ã§ãåççã«ã¯ããããä»çµã¿ãå
¥ãããã¨ã§å¹»è¦ãæ¸ããããã¨ãããã¨ãæ¸ããã®ã ãã©ãå®éã«ãããªæãã§ç²¾åº¦ãããããã¨ãããã¦ã¾ããã
ChatGPTは返答の全体をイメージして答えをはじめる、そして誤っても訂正ができず幻覚を見る - きしだのHatena
å
çãå
¥ãã¦çæããªãããã¨ã§ããããè¿çãã§ããã¨ãã話ã
GPT-4以上? 自分で何度も“推敲”し完成度を上げる言語生成AI「Self-Refine」:Innovative Tech - ITmedia NEWS
â» è¿½è¨ 2023/4/20 ChatGPTèªä½ã§ChatGPTãå¼·åããã°ããã®ã§ã¯ãã¨ãã話ãåºã¦ãã¾ãããGPTãããããããããä»çµã¿ã§ããçæããæç« ãèªåã§å¤å®ãã¦ããã£ã½ãæç« ã«ãªãããã«ãã©ã¡ã¼ã¿ã調æ´ãã¦ããã¾ãã
ã¨ãããã¨ã§ããã°ãã人éãè¶
ãããããªåçãªé²åã¯ãªãããã§ãã©ã使ã£ã¦ãããã£ã¦ãªãããã
â» 2023/4/26 ã·ã³ã®ã¥ã©ãªãã£ããã¦ã人éã«ã¯èªç¥ã§ããªãã®ã§ã¯ãã¨ããããã³ããããã©ãèªç¥ã§ãããã©ããã¯åé¡ã«ãªã£ã¦ãªãã®ã§ããã¾ãæ¬è³ªçã§ã¯ãªãã§ãããã