
ç§åŠççºèŠã®ã¹ã±ãŒãªã³ã°åïŒASI-ARCHã瀺ãAIç ç©¶ã®æ°ããªãã©ãã€ã
ã人éã®ç ç©¶è ã2000æéãããŠè¡ãç ç©¶ããAIã¯ãã£ãæ°æéã§å®äºããããããã¯èªåŒµã§ã¯ãªããå®éã«èµ·ããåºæ¥äºã§ããASI-ARCHãšåŒã°ãã驿°çãªAIã·ã¹ãã ãã人é¡ã®ç§åŠç ç©¶ã«æ°ããª1ããŒãžãå»ã¿ãŸããã
ããã®AIã¯ã誰ãããæç€ºãåããããšãªããèªã仮説ãç«ãŠãå®éšãèšèšããçµæãåæããŠãããã«æ°ãã仮説ãçã¿åºããšãããå®å šã«èªåŸçãªç 究掻åãå®çŸããã®ã§ãããã®ææã¯é©ãã¹ããã®ã§ã106åãã®æ°ããAIæè¡ãçºèŠã
ãäžã«ã¯ãäžçäžã®ç ç©¶è ãã¡ãé·å¹Žè§£æ±ºã§ããªãã£ãåé¡ãžã®ããšã¬ã¬ã³ããªè§£æ±ºçãå«ãŸããŠããŸããã
Spotifyã§ããããããé³å£°é ä¿¡ïŒããã¿ã®AIããã¯ã©ãžãªã
ãäžççµæžãã©ãŒã©ã ã®èª¿æ»ã§ã¯ã2030幎ã«å¿
èŠãªã¹ãã«ãã®ç¬¬5äœã«ã奜å¥å¿ãšç涯åŠç¿ããã©ã³ã¯ã€ã³ã奜å¥å¿ã¯è¶£å³ã§ã¯ãªãçµæžçå¿
é ã¹ãã«ã
æèãAIæä»£ã®æåŒ·ã¹ãã«ã奜å¥å¿åããã§è©³èª¬ããŠããŸãã
ãŸããã
ãçŸä»£ç€ŸäŒã«ãããŠãAIã®èœåã¯ææ°é¢æ°çãªæé·ãéããŠããŸãããã®äžæ¹ã§ãAIç ç©¶ãã®ãã®ã®é²æ©ã¯ãäŸç¶ãšããŠäººéã®ç ç©¶è ã®èªç¥èœåãšããç·åœ¢çãªå¶çŽã«çžãããŠããŸããããã¯ãAIéçºã«ãããæ·±å»ãªããã«ããã¯ãçã¿åºããŠããæ ¹æºçãªãã©ããã¯ã¹ã§ãã
ããã®äººéäžå¿ã®éçºã¢ãã«ã§ã¯ãã€ãããŒã·ã§ã³ã®é床ã¯èšç®èœåã§ã¯ãªãã人éã®ç ç©¶è ã®æèãäœæ¥ã®åž¯åå¹ ã«ãã£ãŠèŠå®ãããŠããŸãã®ã§ãã
ããã®èª²é¡ãä¹ãè¶ãããããäžã€ã®å€é©çãªããžã§ã³ãæå±ãããŠããŸããããã¯ãAIãèªåŸçã«ç§åŠç ç©¶ãè¡ãããã匷åãªæ¬¡äžä»£ã¢ãã«ãèšèšãããAIç ç©¶ã®ããã®äººå·¥è¶ ç¥èœïŒASI4AIïŒããšããæ§æ³ã§ãã
ãç§åŠã®æŽå²ãæ¯ãè¿ããšãæã ã¯ãããŸã§äžã€ã®äž»èŠãªãã©ãã€ã ãçµãŠããŸããã第äžã¯ã芳枬ãšå®éšã«åºã¥ããçµéšçç§åŠãã第äºã¯ãæ³åãæ¹çšåŒã§äžçãèšè¿°ãããçè«çç§åŠãããããŠç¬¬äžããã³ã³ãã¥ãŒã¿ã·ãã¥ã¬ãŒã·ã§ã³ãé§äœ¿ãããèšç®ç§åŠãã§ãã
ããããŠä»ãæã ã¯ãã®æ¬¡ã®æ®µéãããªãã¡ãAIé§ååã®ç§åŠçºèŠããšãã第åã®ãã©ãã€ã ã®å¹éãã«ç«ã¡äŒã£ãŠããã®ãããããŸãããããã¯åã«ç ç©¶ãé«éåããã ãã§ãªããAIã仮説çæãåéæšªæçãªç¥ã®çµ±åã«ãããŠãçã®ååè ãšãªãå¯èœæ§ã瀺åããŠããŸãã
ãæ¬çš¿ã§è©³è¿°ãããASI-ARCHãã¯ããã®ç¬¬åã®ãã©ãã€ã ãäœçŸããããŸãã«äž»åœ¹ãšåŒã¶ã¹ãååšã§ããAIç ç©¶ã®äžã§ãç¹ã«éèŠãã€å°é£ãªé åã§ãããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ã¢ãŒããã¯ãã£çºèŠã«ãããŠãASI-ARCHã¯AIãèªåŸçã«ç§åŠç ç©¶ãè¡ãåã®ãã¢ã³ã¹ãã¬ãŒã·ã§ã³ãšããŠç»å ŽããŸããã
ãããã¯ãAIãåãªããéå ·ããããç¥ã®æ¢æ±ã«ããããååè ããžãšæè¯ããã詊ã¿ã§ããæ¬çš¿ã§ã¯ããã®ç»æçãªã·ã¹ãã ã®æ§é ãè§£åãããã®æè¡çèæ¯ãšææããã®ç€ŸäŒçã€ã³ãã¯ãã«ã€ããŠèå¯ããŠãããŸãã
èæ¯ïŒAIç ç©¶ã®ããã«ããã¯
ãAIã·ã¹ãã ã®æ§èœã¯ææ°é¢æ°çã«åäžããŠããŸãããããã®ç ç©¶éçºã®é床ã¯äººéã®èªç¥èœåã«å¶çŽãããçŽç·çãªããŒã¹ã«çãŸã£ãŠãããšææãããŠããŸãã
ãææ°ã¢ãã«ãäœãåºãã«ã¯é«åºŠãªå°éç¥èãšå€å€§ãªæéãå¿ èŠã§ãèšç®è³æºã®è±å¯ãããã人éç ç©¶è ã®æãè¶³ããªãããšãããã«ããã¯ã«ãªãã€ã€ãããŸãããã®ãžã¬ã³ãã解決ãããããAIèªèº«ã«ç ç©¶ãä»»ããçºæ³ã泚ç®ãããŠããŸãã
ãäŸãã°ãã¥ãŒã©ã«ã»ã¢ãŒããã¯ãã£ã»ãµãŒãïŒNASïŒãšåŒã°ããææ³ã§ã¯ãã³ã³ãã¥ãŒã¿ãã¢ãã«æ§é ãèªåæ¢çŽ¢ããŸãããããåŸæ¥ã®NASã¯äººéãçšæããéšåã®çµã¿åãããæé©åããã«çãŸããçã«æ¬æ°ãªçºæ³ã¯äººéé Œã¿ã§ããã
ãããã§çŸããã®ãã人工ç¥èœã«ãã人工ç¥èœç ç©¶ïŒAI for AI ResearchïŒã§ããããã¯AIãèªåŸçã«ä»®èª¬ãç«ãŠãå®éšãè¡ããçµæãåæããŠæ¬¡ã®ç ç©¶ã«æŽ»ãããšããç ç©¶ãµã€ã¯ã«ãã®ãã®ãèªååãã詊ã¿ã§ãã
ãè¿å¹ŽãLLMã®é£èºçãªæ§èœåäžã«ãããAIãããã°ã©ã ãæžãããæç« ãçè§£ãããããèœåãå€§å¹ ã«é«ãŸããŸããããã®ãããã§ãAIãç ç©¶è ããšã³ãžãã¢ãšããŠæ¯ãèããç§åŠççºèŠã«å¯äžã§ããå¯èœæ§ãçŸå®å³ã垯ã³ãŠããŸãã
ããã®æèã§èªçããASI-ARCHã¯ãAIãAIã¢ãã«ãèšèšãããšããææŠãæ¬æ Œçã«å®èšŒããåããŠã®ã·ã¹ãã ãšèšããŸããç ç©¶ããŒã ã¯ãã®ææããã¢ãã«ã¢ãŒããã¯ãã£çºèŠã«ãããAlphaGoã¢ãŒã¡ã³ãããšè¡šçŸããŸãããããã¯ã€ãŸãã人éã«ã¯æãã€ããªãã£ãçŸããã¢ãã«èšèšäžã®äžæãAIã瀺ããç¬éã ããšããæå³åãã§ãã
AIã«ããèªåŸåç ç©¶ã·ã¹ãã ã®æŠèŠ
ãASI-ARCHã¯Artificial Superintelligence for AI Researchã®ç¥ã§ãçŽèš³ããã°ãAIç ç©¶ã®ããã®äººå·¥è¶ ç¥èœããšãã壮倧ãªååã§ãããã®å®æ ã¯ãLLMãå«ãè€æ°ã®AIãšãŒãžã§ã³ããå調ããŠãæ°ãããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ã¢ãŒããã¯ãã£ïŒæ§é ïŒãèªåçã«çã¿åºãè©äŸ¡ãããèªåŸåç ç©¶ãã©ãããã©ãŒã ã§ãã
ãASI-ARCHã¯äžé£ã®ç ç©¶ããã»ã¹ã人æãä»ããã«åãããèšèšãããŠããŸãããã®ãµã€ã¯ã«ã¯å€§ãã3ã€ã®åœ¹å²ïŒã¢ãžã¥ãŒã«ïŒããæãç«ã¡ãŸãã
Researcher
ãæ°ããã¢ãã«ã®ã¢ã€ãã¢ãææ¡ãã詊äœã³ãŒããæžãèµ·ãã圹å²ãéå»ã®å®éšããŒã¿ãèç©ç¥èãåç
§ãã€ã€ãåŸæ¥ã«ãªãçºæ³ã§ã¢ãã«æ§é ïŒãããã¯ãŒã¯ã®å±€æ§æãæ¥ç¶æ¹åŒãªã©ïŒãèæ¡ããŸãã
Engineer
ãææ¡ãããã¢ãã«ãå®éã«åŠç¿ã»è©äŸ¡ãã圹å²ãã³ãŒããå®è¡ããã¢ãã«ãèšç·ŽããŠæ§èœã枬å®ããŸããäžå
·åãããã°ãããã°ããŠä¿®æ£ãè¡ããŸãã
Analyst
ãå®éšçµæãç·åè©äŸ¡ãã圹å²ãã¢ãã«ã®æ§èœææšã ãã§ãªããæ§é äžã®ç¹åŸŽãå®éšãã°ãè§£æããæ¬¡ã®ãµã€ã¯ã«ã«åããæ¹åç¹ãæŽå¯ãå°ãåºããŸãã
ãããã«ASI-ARCHã«ã¯ããããã®ãšãŒãžã§ã³ããã¡ãæ¯ããç¥èããŒã¹ãšããŒã¿ããŒã¹ããããŸããç¥èããŒã¹ã«ã¯ç·åœ¢ã¢ãã³ã·ã§ã³ãªã©é¢é£åéã®éèŠãªè«æç¥èŠããŸãšããããŠãããå¿ èŠã«å¿ããŠResearcherãAnalystãåç §ããŸãã
ãããŒã¿ããŒã¹ã«ã¯ãããŸã§ã«è©Šããã¢ãã«æ¡ããã®çµæãèç©ããããšãŒãžã§ã³ããã¡ã¯éå»ã®æåã»å€±æããåŠç¿ããŸãã
ããã®ããã«ASI-ARCHã¯ãäžã€ã®çµ±åã·ã¹ãã ã®äžã§ææ¡âå®éšâåæã®ã«ãŒããèªååãããã®ã§ããå³4ã«ASI-ARCHã®å šäœæ§æã瀺ããŸãã

å³4ïŒASI-ARCHã®èªåŸç ç©¶ãã¬ãŒã ã¯ãŒã¯æŠèŠã ç ç©¶ãµã€ã¯ã«ã¯çŽ«ã®ãResearcherããæ©ã®ãEngineerããéã®ãAnalystãããããŠç¥èäŸçµŠæºã§ããèµ€ã®ãCognition Baseãã飿ºããŠéã«ãŒãã§é²è¡ãããåãšãŒãžã§ã³ããé ã«æ°èŠã¢ãŒããã¯ãã£ã®ææ¡âèšç·Žã»è©äŸ¡âçµæåæãè¡ãããã®æŽå¯ãæ¬¡ã®ææ¡ã«åæ ãããããšã§AIèªããç¶ç¶çã«ã¢ãã«ãæ¹è¯ããã
æè¡çææ³ã®ãã€ã³ãïŒèªååããããåµé ããšãéžæã
ãASI-ARCHæå€§ã®ç¹åŸŽã¯ãæ¢çŽ¢ç©ºéã®åºããšæ¢çŽ¢ããã»ã¹ã®å·¥å€«ã«ãããŸããåŸæ¥ã®NASãéãããéšåã®çµã¿åããæé©åã ã£ãã®ã«å¯ŸããASI-ARCHã®Researcherã¯éå»ã®ããŒã¿ãèžãŸãã€ã€ã人éãäºæ³ããªãæ¬æ°ãªæ§é ãèªç±ã«çºæ³ã§ããŸããããã¯LLMã®åŒ·åãªæç« çæã»ããã°ã©ãã³ã°èœåãæŽ»çšããâæªç¥ã®ã¢ã€ãã¢âãã³ãŒããšããŠå ·äœåã§ããããã§ãã
ããã£ãšããäœã§ãéé²ã«è©Šãã°è¯ãçµæãåŸãããããã§ã¯ãããŸãããASI-ARCHã§ã¯å¹æçãªæ¢çŽ¢ã®ãããããã€ãã®æŠç¥ãåãããŠããŸãã
ãšãªãŒãããŒã¹ã®é²åæŠç¥
ãããŒã¿ããŒã¹äžã®æ§èœäžäœã®ã¢ãã«ãã芪ããšããŠéžã³ãããããçæ³ãåŸãŠæ°ããå€ç°ïŒããªãšãŒã·ã§ã³ïŒãææ¡ããŸãããã ãäžäœã ãã§ãªãäžå®ç¯å²ã®ã¢ãã«ã«ã©ã³ãã æ§ãæãããŠéžã¶ããšã§ã倿§æ§ã確ä¿ããŠããŸãã
ãã£ãããã¹é¢æ°ã«ããè©äŸ¡
ãææ¡ãããã¢ãã«ã¯è€åçãªè©äŸ¡ææšïŒãã£ãããã¹ïŒã§ã¹ã³ã¢åãããŸããå
·äœçã«ã¯ã(1)åŠç¿æã®æå€±ããã³ãããŒã¯ãã¹ãã§ã®ã¹ã³ã¢ãšãã£ãå®éçæ§èœã®åäžåºŠã(2)ã¢ãã«æ§é ã®æ°èŠæ§ã»åŠ¥åœæ§ã»è€éããªã©å®æ§çè©äŸ¡ãLLMãå°éå®¶ã®ç®ç·ã§å€æããã¹ã³ã¢ãã®2èŠçŽ ãçµã¿åãããŠããŸãã
ãæ°å€ã ãã«é Œããšãææšãããããã«æé©åããŠæ¬è³ªçã§ãªãè§£ïŒãããããå ±é ¬ãããã³ã°ãïŒã«é¥ãæããããããã人éã®å¯©çŸçŒã«è¿ãè©äŸ¡ãåãå ¥ããŠããç¹ãéèŠã§ãã
èªå·±æ¹ååãããã°
ãEngineerã¢ãžã¥ãŒã«ã§ã¯ãã¢ãã«ã®åŠç¿äžã«ãšã©ãŒãç°åžžæ€åºãããã°èªåã§ãã°ãè§£æããResearcherã«ãã£ãŒãããã¯ããŠã³ãŒããä¿®æ£ãããŸããããã«ãããææãªã¢ã€ãã¢ãã³ãŒãäžã®äºçްãªãã°ã§æšãŠãããã®ãé²ããæåŸãŸã§è²ãŠãã工倫ããªãããŠããŸãã
ããŸããæ¢çŽ¢ãå¹çåããããäºæ®µéã®è©äŸ¡ããã»ã¹ãå°å ¥ãããŠããŸããæåã¯å°èŠæš¡ã¢ãã«ïŒäŸãã°ãã©ã¡ãŒã¿æ°æ°åäžçšåºŠïŒã§å€æ°ã®åè£ããã£ãšè©Šããææãªãã®ãéžå¥ããŸãã
ãæ¬¡ã«éžã°ããã¢ãã«ãå€§èŠæš¡ã¢ãã«ïŒäŸãã°æ°åãã©ã¡ãŒã¿ïŒã«ã¹ã±ãŒã«ã¢ããããŠæ¹ããŠæ€èšŒããããšã§ãèšç®è³æºãç¡é§ã«ããä¿¡é Œæ§ã®é«ãè©äŸ¡ãå¯èœã«ããŠããŸãã
AIãçºèŠããæ°ã¢ãŒããã¯ãã£ã®æ§èœãšç¹åŸŽ
ãASI-ARCHã¯çŽ20,000 GPUæéã«ããã1,773ãã®å®éšãèªåŸçã«è¡ãã106çš®ãã®é«æ§èœãªãã¥ãŒã©ã«ãããã¯ãŒã¯ã»ã¢ãŒããã¯ãã£ãçºèŠããŸããã
ããããã¯ãã¹ãŠç·åœ¢æéã®ã¢ãã³ã·ã§ã³æ©æ§ïŒTransformerã®é«éåããªãšãŒã·ã§ã³ïŒã®ã¢ãã«ã§ãåŸæ¥ã®äººéèšèšã®ããŒã¹ã©ã€ã³ãäžåãæ§èœã瀺ãããã®ã§ããç¹ã«5ã€ã®ä»£è¡šçãªæ°ã¢ãã«ã«ã€ããŠãããŒã¹ã©ã€ã³ææ³ãšæ¯èŒããæ§èœã衚1ã«ç€ºããŸãã
衚1ïŒèšèªã¢ããªã³ã°ããã³åžžèæšè«ãã³ãããŒã¯ã§ã®ã¢ãã«æ§èœæ¯èŒã
ã人éãèšèšããããŒã¹ã©ã€ã³ã¢ãã«ïŒäž3è¡ïŒãšãASI-ARCHãèªåŸçã«çºèŠããã¢ãã«ïŒäž5è¡ïŒãšã®æ¯èŒãååã¯èšèªã¢ããªã³ã°æ§èœïŒPerplexityãäœãã»ã©è¯ãïŒãéžæåé¡ã®æ£è§£ç(%)ã§ã倪åã¯æé«å€ã
ãASI-ARCHç£ã®ã¢ãã«çŸ€ã¯å°èŠæš¡ïŒçŽ3å4åäžãã©ã¡ãŒã¿ïŒãªãããã»ãŒå šãŠã®ææšã§ããŒã¹ã©ã€ã³ãäžåã£ãŠãããå¹³åçãªæ§èœææšã§ãããŒã¹ã©ã€ã³ãæç¢ºã«åé§ããŠããã

ã衚1ãããASI-ARCHãçºèŠããã¢ãã«ïŒè¡šäžã§ã¯éãããããã®ã¢ã€ã³ã³ã§ç€ºãããŠããŸãïŒããåŸæ¥ã®äººéèšèšã¢ãã«ïŒãã³ã¯ã®äººç©ã¢ã€ã³ã³ïŒããåºç¯ãªã¿ã¹ã¯ã§è¯å¥œãªæçžŸãåããŠããããšãåãããŸãã
ãäŸãã°PathGateFusionNetãContentSharpRouterãšãã£ãæ°ã¢ãã«ã¯ãèšèªã¢ãã«ã®å°é£ãªèªè§£ã¿ã¹ã¯(LAMBADA)ãåžžèæšè«(HellaSwagãªã©)ã§ããŒã¹ã©ã€ã³ãäžåã粟床ã瀺ããŠããŸãã
ãæ³šç®ãã¹ãã¯ããããã®æ§èœåäžãåã«ç¹å®ã¿ã¹ã¯ã ãã§ãªãã平忧èœã®åºäžããšããŠçŸããŠããç¹ã§ããããªãã¡ãASI-ARCHã¯æ±çšçã«åªããã¢ãã«ãäœãåºãããšèšããŸãã
ãã§ã¯ASI-ARCHã¯ã©ã®ããã«ããŠãããåªããã¢ãŒããã¯ãã£ãèŠåºããã®ã§ãããããè§£æã®çµæãè峿·±ãèšèšãã¿ãŒã³ã®åŸåãæããã«ãªã£ãŠããŸããéçºãããã¢ãã«ã®å€ãã¯ã人éã®èšèšã§ã¯åå¥ã«ç ç©¶ãããŠããææ³ãå·§ã¿ã«çµã¿åããããã€ããªããæ§é ãåããŠããŸããã
ãäŸãã°ã屿çãªç³ã¿èŸŒã¿åŠçããšãé·è·é¢äŸåãå¶åŸ¡ããã²ãŒãã£ã³ã°æ©æ§ããåæã«æã€ãããã¯ãŒã¯ããç°ãªãæç³»åã¢ããªã³ã°ææ³ãéå±€çã«çµã¿åãããæ§é ãªã©ã人éãäžåºŠã«æãã€ãã®ãé£ããè€åçãªã¢ã€ãã¢ã倿°å«ãŸããŠããã®ã§ãã
ããã®ããšã¯ãASI-ARCHãæ¢åç¥èŠã®åçºæã«çãŸãããç°åéã®ç¥èãèåããåµçºçãªãã¶ã€ã³ã«å°éããããšã瀺åããŠããŸãã
ãããã«ãASI-ARCHã®æ¢çŽ¢éçšãã¢ãã¿ãŒããããšã§ãæéçµéã«äŒŽãæ§èœåäžã®æ§åã確èªãããŠããŸããå³6ã«ãæ¢çŽ¢ã®é²è¡ãšãšãã«ã¢ãã«æ§èœãã©ãå€åãããã瀺ããŸãã

å³6ïŒASI-ARCHã®æ¢çŽ¢ã«ãããæ§èœææšã®æšç§»ã
ã(a) äžäœåè£ã¢ãã«ã®å¹³åãã³ãããŒã¯ã¹ã³ã¢ïŒçŽ«ïŒãšæå€±å€ïŒæ©ïŒã®æšç§»ãåææ®µéã§æ§èœãæ¥éã«åäžãããã®åŸãç·©ããã§ã¯ãããçå®ã«æ¹åããŠããã
ã(b) ãã£ãããã¹ã¹ã³ã¢ïŒè€åè©äŸ¡ïŒã®å¹³åå€ïŒéïŒã®æšç§»ãåŠç¿ãé²ãã«ã€ãã¹ã³ã¢ã¯é æã¡ã«ãªã£ãŠããããããã¯æ¥µç«¯ãªå€ãæããè©äŸ¡é¢æ°ã®ç¹æ§ã«ãããã®ã§ãå®éã®çããŒã¿ææš(a)ã¯æ¢çŽ¢ãé²ãã»ã©åäžãç¶ããŠããç¹ã«æ³šç®ã
ãå³6(a)ãããASI-ARCHã¯æ¢çŽ¢ãµã³ãã«æ°ïŒè©Šè¡ããã¢ãã«æ°ïŒã®å¢å ã«äŒŽã£ãŠããã³ãããŒã¯æ§èœãå³è©äžããã«æ¹åããæå€±ïŒèª€å·®ïŒã¯äžè²«ããŠäœäžããŠããŸããåºç€ããæ¥æ¿ãªè¯åãèŠããããã®åŸã¯ç·©ãããªäŒžã³ã«ç§»è¡ããŸãããæåŸãŸã§åäžåŸåãç¶ããŠããç¹ã§é æã¡ã«ã¯ãªã£ãŠããŸããã
ãäžæ¹ãå³6(b)ã®è€åã¹ã³ã¢ã¯åŸåã§æšªã°ãã«ãªããŸãããããã¯åè¿°ã®ããã«ãã£ãããã¹è©äŸ¡ãã·ã°ã¢ã€ã倿ã«ãã£ãŠé£œåããèšèšã®ããã§ãã
ãèŠããã«ãASI-ARCHã¯ååãªèšç®è³æºãäžããã°äžããã»ã©çå®ã«é«æ§èœã¢ãã«ãèŠåºãç¶ããããšã瀺åãããã®ã§ãããã®çŽç·çãªé¢ä¿ã¯ãç ç©¶ããŒã ã«ãã£ãŠãç§åŠççºèŠã®ã¹ã±ãŒãªã³ã°åããšãåŒã°ããŠããŸãã

ãå®éãå³1ã«ã¯ãæ¶è²»èšç®æéãšçºèŠããæåã¢ãŒããã¯ãã£ã®çŽ¯ç©æ°ããããããããŠãããã»ãŒæ¯äŸé¢ä¿ã«ããããšã瀺ãããŠããŸããããã¯ãç ç©¶éçºã®é²æ©ã人æã§ã¯ãªãèšç®è³æºã«ããã¹ã±ãŒã«å¯èœã§ãããããšãæå³ããç»æçãªç€ºåãšèšããã§ãããã
èå¯
ãASI-ARCHããããããæå€§ã®è¡æã¯ãAIãèªåŸçã«ç¥çåµé 掻åãè¡ããããšãå®èšŒããç¹ã§ãã人éãæ°ã«æãããŠè©Šè¡é¯èª€ãããããªã¢ãã«æ¹è¯ã®ããã»ã¹ããAIã¯èšå€§ãªèšç®åãæŠåšã«çæéã§æãéããŸããã
ãç¹ã«è峿·±ãã®ã¯ãæ°ã¢ãã«ã®èšèšã«ãããŠéå»ã®äººéã®ç¥èŠä»¥äžã«ãèªèº«ã®å®éšããåŸãæŽå¯ã掻çšããããšããåæçµæã§ãã
ãããã¯ã人éãçšæããæç®ç¥èããããAIèªèº«ãçã¿åºããçµéšç¥ã«éã¿ã眮ããŠæ¬¡ã®ã¢ã€ãã¢ãèæ¡ããããšã瀺åããŸããããã°AIãç¬åŠã§æ°ããªåçãçºèŠãããšãèšããçŸè±¡ã§ãããAIç ç©¶ã®èªåŸæ§ãäžæ®µãšæ·±ãŸã£ãããšãæå³ããŸãã
ããã£ãšãããã®ç ç©¶ã¯å§ãŸã£ãã°ããã§ãããããã€ãã®èª²é¡ãçæç¹ãææãããŠããŸãã
èšç®è³æºã®å€§éæ¶è²»
ãä»åASI-ARCHã¯çŽ20,000 GPUæéãã®èšç®æéãè²»ãããŸãããããã¯å€§åŠãç ç©¶æ©é¢ã§äœ¿ãããªãœãŒã¹ãšããŠã¯æ¥µããŠå€§ãããå®éšèŠæš¡ãããã«æ¡å€§ããã«ã¯è²»çšé¢ã»ç°å¢é¢ã®èª²é¡ããããŸããããå¹ççãªæ¢çŽ¢ã¢ã«ãŽãªãºã éçºããèšç®è³æºãå
±æããä»çµã¿äœããæãŸããŸãã
æ±çšæ§ãšåçŸæ§ã®æ€èšŒ
ãæ¬ç ç©¶ã¯ç·åœ¢ã¢ãã³ã·ã§ã³ãšããç¹å®åéã§ã®æåäºäŸã§ããä»ã®åéïŒäŸãã°ç»åã¢ãã«ã®æ§é èšèšãæé©åã¢ã«ãŽãªãºã ã®çºèŠãªã©ïŒã«ãåæ§ã®ã¢ãããŒããéçšããã®ããä»åŸã®æ€èšŒãå¿
èŠã§ãããŸããä»åã®ææãã®ãã®ããã¢ã¬ãã¥ãŒåã®ãã¬ããªã³ã段éã§ãããã³ãã¥ããã£ã«ããç¬ç«æ€èšŒãåçŸå®éšãçµãŠç¢ºããªãã®ãšãªãã§ãããã
人éãšã®å調ãšå¶åŸ¡
ãAIãèªåŸçã«ç ç©¶éçºãé²ãããããšããããšã¯ãè£ãè¿ãã°äººéã®çè§£ãè¶
ããã¢ãã«ãç»å ŽãåŸãããšãæå³ããŸããASI-ARCHãææ¡ããã¢ãã«ã®äžã«ãããã®åäœåçãçŽèгçã«æŽã¿ã«ãããã®ããããããããŸããã
ã人éã®ç ç©¶è ã¯ãAIãçã¿åºãããã®ãé©åã«è©äŸ¡ã»éžå¥ãã瀟äŒã«æçãªåœ¢ã§æŽ»çšããŠãã責任ããããŸãããŸããASI-ARCHèªäœã®ç®ç颿°èšèšãå®å šæ§ãéèŠã§ããæªæããç®çã§èªåŸç ç©¶AIãå©çšãããªããããå«çé¢ã»ã¬ããã³ã¹é¢ã§ã®è°è«ãä»åŸæ·±ããå¿ èŠãããã§ãããã
é¢é£æ å ±ã»åèæç®
Transformerã®åç¹ãç¥ãããã«
Vaswani, A. et al. (2017). "Attention is All You Need."
仿¥ã®çæAIã®åºç€ãç¯ããTransformerã¢ãŒããã¯ãã£ãææ¡ãããèšå¿µç¢çãªè«æã§ãã
å¹ççãªã¢ãŒããã¯ãã£ã®æœ®æµãçè§£ããããã«
Gu, A., & Dao, T. (2023). "Mamba: Linear-Time Sequence Modeling with Selective State Spaces."
Transformerã®æåãªå¯ŸæéЬã§ããMambaã¢ãŒããã¯ãã£ã«é¢ããäž»èŠãªè«æãASI-ARCHãæãã ç ç©¶é åã®åºãããçè§£ããå©ãã«ãªããŸãã
èªå·±æ¹åAIãšå®å
šæ§ã«ã€ããŠæçŽ¢ãæ·±ããããã«
Bostrom, N. (2014). "Superintelligence: Paths, Dangers, Strategies."
é«åºŠãªAIãããããé·æçãªèª²é¡ã«ã€ããŠèå¯ããå€å
žçåèããããã¯ãããæè¿ã®AIã¢ã©ã€ã¡ã³ãã«é¢ããåŠè¡è«æ ããæ¬çš¿ã§è§Šããå«ççèŠåº§ãæ·±ããäžã§æçã§ãã Â
ããšãã
ãAIãèªãã®åŸç¶ãšãªãAIãçã¿åºãããã€ãŠã¯ç©ºæ³ã«éããªãã£ããã®ã·ããªãªããASI-ARCHã®ç»å Žã§çŸå®å³ã垯ã³ãŠããŸããããã¡ãããçŸæç¹ã§äººéã®ç ç©¶è ãäžèŠã«ãªãããã§ã¯ãããŸããã
ãããããç ç©¶è ã®åœ¹å²ãå€ããã€ã€ããã®ã¯ç¢ºãã§ãããããŸã§ã¯äººéã詊è¡é¯èª€ããŠããéšåãAIãæ ãã人éã¯ããé«ãèŠç¹ããæ¹åæ§ã瀺ããããå«çã»ç€ŸäŒçãªèгç¹ãçµ±åããããã圹å²ãžãšã·ããããŠãããããããŸããã
ãASI-ARCHã®ææã¯ãAIãæŽãªãAIãèšèšããããšããèªå·±èšåçãªé²æ©ã®ç¬¬äžæ©ã§ãããã®ã€ã³ãã¯ãã¯æè¡é¢ã«çãŸãããç§ãã¡ãåµé æ§ãç¥èœãæããèãæ¹ã«ãå€é©ãè¿«ãã§ãããã
ãæ¬çš¿ãéããŠåŸãããã¢ã€ãã¢ãç¥èããããªãã®ããžãã¹ã«å°ãã§ã圹ç«ã€ããšãé¡ã£ãŠããŸãããããæ¬çš¿ãåèã«ãªã£ããšæããŠããã ããŸãããããã²ããããããããã©ããŒããããŠããã ãããšå±ã¿ã«ãªããŸããä»åŸãå®è·µçãªããŠããŠãAIææ°ååãå ±æããŠãããŸãã®ã§ãåŒãç¶ããèªã¿ããã ãããšå¬ããã§ãã
â ïžãéèŠïŒChatGPTã䜿ãããªãã人ã¯å
šäœã®3å²ä»¥äž
ããªããïŒã質ååãã®å·®ã§ãã
ãå®ã¯ãAIãã10åã®äŸ¡å€ãåŒãåºããåãã®ç«ãŠæ¹ãã«ã¯ç§åŠçãªæ³åããããŸããGoogleã20%ã«ãŒã«ã§å®èšŒããAmazonã®ããŸã¹ãæšå¥šããæ¹æ³ãå
šãŠãAIæä»£ã®æåŒ·ã¹ãã«ã奜å¥å¿åããã§å
¬éäžã
ãAIãç¬æã«ãçãããè¿ãä»ãåè² ã®åããç®ã¯ãåããã«ããããã®ç§å¯ã𿿥ã®ä»äºã§äœ¿ãã13ã®å®è·µããŒã«ãåçž®ããæèãAIæä»£ã®æåŒ·ã¹ãã«ã奜å¥å¿åããããã²ãã§ãã¯ããŠã¿ãŠãã ããã
ãGoogleã®ããã¯ãªãŒããTom Longæ°ãæå±ãããè¯ãã³ãŒã vs æªãã³ãŒããã®å¯Ÿæ¯åœ¢åŒã§åŠã¶å®è·µã¬ã€ãããåãã³ãŒãããããè¯ãã³ãŒãããžãé·æçã«ä¿å®ã»æ¡åŒµå¯èœãªã³ãŒããæžãããã®å®è·µçãªæéãäœç³»åããå¿ èªæžã§ãã
ãåã«åäœããã ãã§ãªããèªã¿ãããããã¹ãããããã倿Žã«åŒ·ãã³ãŒããšã¯äœãïŒæ¬æžã§ã¯ãåœåèŠåã颿°èšèšããšã©ãŒåŠçãæœè±¡åã®ååãªã©ãããã®ãšã³ãžãã¢ãæ¥ã å®è·µããŠãããè¯ãã³ãŒããã®æžãæ¹ããè±å¯ãªå®äŸãšãšãã«è§£èª¬ããŸãã
ãç¹çãã¹ãã¯ãæè¡çãªæ£ããã ãã§ãªããããŒã éçºãã³ãŒãã¬ãã¥ãŒã§ã®èгç¹ãç¶²çŸ ããŠããç¹ãã³ãŒãã®å¯èªæ§ãä¿å®æ§ãæ¡åŒµæ§ãé«ããããšã§ãéçºé床ã®åäžãšãã°ã®åæžãå®çŸã§ããŸããç¶ç¶çãªå質åäžãç®æããšã³ãžãã¢ãéçºãªãŒããŒã«æé©ãªäžåã§ãã
ããžã¿ã«äžçã®ããªãïŒããããããçŸä»£äººã®å¿
èªæž
ãããªã³ã¹ãã³å€§åŠã®äŒèª¬çææãã©ã€ã¢ã³ã»ã«ãŒããã³ããã³ã³ãã¥ãŒã¿ãŒãµã€ãšã³ã¹ã®åºç€ãå°éçšèªãé¿ããªããäžå¯§ã«è§£èª¬ãããã°ã©ãã³ã°çµéšããªããŠããã€ã³ã¿ãŒããããããŒããŠã§ã¢ããœãããŠã§ã¢ãAIãã»ãã¥ãªãã£ãŸã§ãæè¡ã®æ žå¿ãçè§£ã§ããŸãã
ã第2çã§ã¯ææ°ãããã¯ãç¶²çŸ ããçŸä»£ç€ŸäŒã§å¿ é ã®ããžã¿ã«ãªãã©ã·ãŒã身ã«ã€ãæ§æã«ãããžãã¹ããŒãœã³ããåŠçãŸã§ãããžã¿ã«äžçã®ä»çµã¿ãæ£ããçè§£ããè³¢ãæŽ»çšããããã®ç¥èãè©°ãŸã£ãäžåã§ãã
ãæ¥ç«32äžäººã®å®è·µããçãŸãããæ¬ç©ã®äŒæ¥åãAI掻çšã¬ã€ãïŒãå®è·µ çæAIã®æç§æžãã¯ã1,000以äžã®å®èšŒæžã¿äºäŸãããšã«ãããžãã¹ã§ç¢ºå®ã«ææãåºãæ¹æ³ã解説ã ç¬èªã®ãé£æåºŠÃ广æ§ãããªã¯ã¹ãã§ãããªãã®çµç¹ã«æé©ãªAIå°å ¥é åºãäžç®çç¶ã
ãæ¥åå¹çåãã·ã¹ãã éçºãã³ãŒã«ã»ã³ã¿ãŒãã€ã³ãã©ä¿å®ãããŒã¿åæã®5åéã§å³å®è·µã§ããæŽ»çšæ³ãåé²ã ããã³ãããšã³ãžãã¢ãªã³ã°ãRAGå®è£ ããããžã§ã¯ã管çãããã»ãã¥ãªãã£ã»ã³ã³ãã©ã€ã¢ã³ã¹å¯ŸçãŸã§ãäŒæ¥ãæ¬åœã«å¿ èŠãšããç¥èãç¶²çŸ ãçè«æžãå人åãã¬ã€ããšã¯äžç·ãç»ããçŸå Žã§äœ¿ããå®è·µæžã§ãã