Iyini i-Fugu ye-Sakana AI, futsi Kungani Wonkhe Muntfu Akhuluma Ngayo?

Kulelisontfo, i-Sakana AI lese-Tokyo — inkampani lensha ye-AI lebaluleke kakhulu e-Japan — yetfula i-Fugu ne-Fugu Ultra, futsi lokwetfulwa kutsatfwa njengelishintjo lelikhulu endleleni i-AI lesetulu lakhiwa ngayo. Esikhundleni sekuceceshwa kwe-model yinye lenkhulu kakhulu kutsi yente konkhe, i-Fugu iyatetfula "njengesistimu yema-agent lamanyenti lelethwa njenge-model yinye" leyo ihlela ngebuhlakani ema-model lamanyenti e-AI lahamba embili ngemuva kwesikhombisi sinye. Kubanikati bemabhizinisi, indzaba lenkhulu akusiwo umncintiswano lomunye wemabhentjimakhi — kodvwa yindlela lensha yekunciphisa kutsembela kwakho kunoma ngumuphi umhlinzeki munye we-AI.

Ngamafushane: i-Fugu yisistimu ye-AI lekhetsa, lehlanganisa futsi lehlela ngekutentekela ema-model lamanyenti lasembili kakhulu ("lichibi layo lema-agent") kute ixazulule umsebenti, bese ihlola umphumela — konkhe ngekusebentisa i-API yinye. Ngobe ema-model langaphansi angashintjaniswa, isistimu ingasusa umsebenti kunoma ngumuphi umhlinzeki lovinjelwako, lohamba kancane noma longatfolakali, futsi loku kucina lokunjalo kanye lebekungekho emabhizinisini.

I-Fugu Isebenta Njani: Kuhlela Esikhundleni Se-model Yinye Lenkhulu

Lunyenti lwemikhicito ye-AI loyisebentisa lamuhla yi-model yinye levela enkampanini yinye. I-Fugu itsatsa indlela lehlukile. Ngekusho kwe-Sakana, ishukunyiswa ngemasistimu lamabili elucwaningo — omabili ashicilelwe ku-ICLR 2026:

  • TRINITY — i-model lencane "yekuhlela" lenika imisebenti kulichibi lema-LLM lahlukahlukene ngekushintjana, ibavumela kutsi basebentisane ngaphandle kwekuhlanganisa ema-weights abo noma kwabelana ngesakhiwo sinye.
  • Conductor — i-model lececeshwe ngekufundza ngemivuzo (reinforcement learning) lehlela kutsi ema-agent akhulumisana njani, futsi ibhalele ngayinye i-model lesebentako ticondziso letentelwe yona.

Umphumela uziphatsa njengelicembu leliholwa kahle kunekuba ngulocwepheshile munye: lohlelako unquma kutsi ngubani lowenta ini, labasebentako baletsa emandla abo, futsi isistimu ihlola imphendvulo ngaphambi kwekuyibuyisela. I-Sakana iyiletsa ngemazinga lamabili — i-Fugu, leyenteleke kutsi ibe ngekushesha emisebentini yemalanga onkhe futsi ifakwe emathulusini ekubhala ikhodi, ne-Fugu Ultra, lehlela emachibi lajulile ema-agent etinkinga letinkhimbinkhimbi letinetinyatselo letinyenti. Inkampani itsi i-Fugu Ultra iyalingana — futsi kulamanye emahlelo iyendlula — ema-model lahamba embili lasembili kakhulu kumabhentjimakhi ebunjiniyela, esayensi nekucabanga.

Indzaba Yeliciniso Yebhizinisi: Kuphunyula Ekubopheni Kumhlinzeki Munye

Enkampanini yase-Saudi Arabia, incenye lebaluleke kakhulu yalokwetfulwa akusilo luhlu lwekulandzelana — kodvwa kukhona kutfutfa. Nangabe umkhicito wakho, lusito lwakho lwemakhasimende noma kusebenta ngekutentekela kwakho kwangekhatsi kuboshelwe kumhlinzeki munye we-AI, utfola tingoti talowo mhlinzeki: kukhuphuka kwemanani kungazelele, mikhawulo yekusebentisa, imikhawulo yekutfolakala ngetifundza neyekulawulwa kwekutfumela emaveni langaphandle, kumiswa kwema-model, nekuphazamiseka kwelusito. Sesake sabhala ngaphambilini ngekutsi kuyekiswa kwe-model yinye kungaphazamisa njani libhizinisi ngebusuku bunye.

Umklamo we-Fugu uphendvula loko ngco. Lichibi layo lema-agent liyashintjaniswa, futsi isistimu "ngendlela lehambisanako iyalicondzisa litrafikhi kute ilidlulise kunoma ngumuphi umhlinzeki lovinjelwe noma lowehlile, kute igcine lusito luchubeka." Ngemagama lalula: nangabe i-model yinye ihamba kancane, ibita kakhulu noma ivinjelwa, umsebenti udlulela kulenye ngaphandle kwekutsi i-application yakho iphuke. Nobe ungayisebentisa ngco i-Fugu noma cha, lesi sakhiwo lekuyiwa kuso yimakethe.

Loku Kusho Kutsini Emabhizinisini ase-Saudi Arabia

I-Saudi Arabia imemetele kutsi 2026 ngumnyaka weBunhlakanipho Lobentiwe, futsi kusetjentiswa kwayo kuyashesha etitolo, kutekutfutsa timphahla, etemphilo nasetinsitakalweni tahulumende. Njengobe tinkampani letinyenti tiyisa i-AI emsebentini wangempela, umbuto uyashintja usuka ku-"ngiyiphi i-model lencono lamuhla?" uye ku-"ngiyigcina njani i-AI yami isebenta ngekutsembeka iminyaka leminyenti?" Kuhlela ema-model lamanyenti ndzawonye kuyimphendvulo lesemcoka:

  • Kucina: akukho ndzawo yinye yekuhluleka nangabe umhlinzeki munye ehlulekile noma ashintja inchubomgomo.
  • Kulawula tindleko: tfumela imisebenti lelula kuma-model langabiti kakhulu futsi ugcinele ema-model laphambili tinkinga letimatima.
  • Kwekhwalithi lencono kakhulu: sebentisa i-model ngayinye lapho inemandla khona esikhundleni sekuphoca i-model yinye kutsi yente konkhe.
  • Kulandzela imitsetfo nekutibusa: likhono lekufaka ema-model esifundza noma asekhaya njengobe i-AI yase-Saudi Arabia letibusako (njenge-ALLaM ne-HUMAIN) ikhula.

Kufanele Usebentise i-Fugu, Noma Umcabango Losemuva Kwayo?

I-Fugu iyatfolakala lamuhla kumakhasimende emabhizinisi ngekusebentisa i-API ye-Sakana (ine-endpoints letijwayelekile te-Chat Completions ne-Responses) futsi ifakwa emathulusini elonjiniyela ngumyalo munye. Nangabe unelicembu lebunjiniyela lelihlola i-AI lesembili kakhulu, ifanele kuhlolwa ngelucwaningo lolusemcoka. Kodvwa awuyidzingi i-Sakana kute uzuze kulesifundvo; umtsetfosimiso — ungalibophi nanini libhizinisi lakho ngco ku-model yinye — ungasetjentiswa kunoma nguyiphi iphrojekthi ye-AI loyicala lamuhla.

Sinyatselo lesiphatsekako kwakha singcimba sekuhlukanisa emkhatsini we-application yakho ne-model, kuze kushintja noma kuhlanganisa bahlinzeki kube lushintjo lwekuhlela kwesistimu, hhayi kubhala kabusha. Emazinga lanjenge-Model Context Protocol (MCP) nema-SDK langabopheki kumhlinzeki munye enta loku kube nekwenteka kusoftwe lejwayelekile yemabhizinisi, hhayi emalabhorethri elucwaningo kuphela.

I-Origami Iyakha Njani i-AI Lengabopheki Kumhlinzeki Munye

Ku-Origami sifaka i-AI emasistimu ebhizinisi angempela — ema-chatbot, basiti bangekhatsi, kusebenta ngekutentekela nema-app lentelwe wena — futsi siphatsa i-model njengencenye lengashintjwa kusukela ngelilanga lekucala. Loko kusho kucondzisa, tindlela tekuwela emuva netikhombisi letikhanyako, kuze kutfutsela kwakho kusinde ekwetfulweni kwe-model lelandzelako, elushintjweni lwemanani noma elushintjweni lwemitsetfo. I-Fugu kucinisekiswa lokubonakala kakhulu kwendlela lesesiyincoma kakade: hlela, ungabopheki. Nangabe ufuna i-AI yakho ihlale isheshisa, ingabiti kakhulu futsi itsembeke njengobe simo sichubeka sishintja, lapho kufanele ucale khona.

Imitfombo

  • Sakana AI — Fugu (lesemtsetfweni): github.com/SakanaAI/fugu
  • Nikkei Asia — umbiko ngekwetfulwa kwe-Fugu nemaphuzu ayo uma icatsaniswa ne-Fable 5 ne-GPT-5.5.
  • VentureBeat — kusebenta kwe-Sakana ezingeni lelisembili kakhulu ngesistimu yema-model lamanyenti ye-Fugu.
  • AI News — kunciphisa kubopheka kumhlinzeki munye nge-Sakana AI Fugu.