Jev Uyasebenza

UJev uyimodeli ye-AI yokwenza izinqumo ezihlelekile eyethulwe yi-TypeSafe AI. Imodeli ayixoxi, ayibhali khodi futhi ayikhiqizi umbhalo; yenza into eyodwa kuphela: izinqumo ezivamile. Ithatha ulwazi olungahlelekile, yenze ukubala okuhambisanayo kanye, bese ikhipha ngokushesha izahlulelo ezihlelekile ezinethuba elilinganisiwe. Ukubambezeleka kungama-70–500 milliseconds kuphela, okokufaka kubiza u-0.042 USD/ama-token ayisigidi, okokukhipha kumahhala, kanti izinga lamaphutha ohlobo lingu-0%. Imodeli ibekwe njengengxenye yokuqala ye-reflex arc yezinhlelo ze-Agent, ithwale imisebenzi efana nokukhetha umzila, ukuhlukanisa, ukuhlolwa kokuqala kobungozi nokuhlunga okuqukethwe, ukuze amamodeli amakhulu athuthukile agxile ekucabangeni okuyinkimbinkimbi, ibe umenzi wezinqumo onolwazi lomhlaba.

Imiphumela yezinqumo ehlelekile: Imodeli ibuyisa kuphela i-Choice (ukukhetha), i-Score (amaphuzu) noma i-Noul (ukwahlulela kwe-boolean) ngaphakathi kwezinhlobo ezichazwe kusengaphambili, ihambisane nethuba elilinganisiwe.
Izahlulelo eziningi ezihambisanayo: Ucingo olulodwa lungaphendula imibuzo eminingi ezimele ngesikhathi esisodwa, futhi ukubambezeleka akukhuli ngokulingana nenani lemibuzo.
Ukuqondisa ngokwamazinga okuzethemba: Imiphumela ihlukaniswa ngokuzenzakalelayo ngokwezinga lokuzethemba elilinganisiwe: ukuzethemba okuphezulu kwenziwa ngokuqondile, okuphakathi kuthunyelwa kumodeli enkulu ukuze kuqinisekiswe, okuphansi kudluliselwe kumuntu.
Isendlalelo sokuhlunga ngaphambi kwe-Agent: Sithatha izinqumo ezincane ezivamile ezifana nokuqondisa umzila, ukuhlukanisa, ukuqinisekisa nokuhlola isimo, sinciphise izingcingo zamamodeli amakhulu abizayo ngaphezu kuka-90%.
Isiqinisekiso samaphutha efomethi anguziro: Imiphumela ikhawulelwe ngokuqinile ku-Schema ebekwe kusengaphambili, ngakho izinga lamaphutha ohlobo lingu-0% ngokwezibalo futhi ingafakwa ngqo emgqeni wokusebenza wekhodi.
Izahlulelo zesikhathi sangempela ezingeni lama-millisecond: Impendulo yemizuzwana engu-70–500 kanye nesinqumo esisodwa esibiza cishe u-$0.0004 kwenza i-AI ingene ezimweni ezidinga ukubambezeleka okuphansi, njengokulawula imidlalo, ukuhlunga okuqukethwe ngesikhathi sangempela nezinqumo zokuhweba.

Kususwa ukukhiqizwa kwe-autoregressive, kusetshenziswe ukusampula okuhambisanayo: Ama-LLM endabuko akhiqiza ngokulandelana i-Token ngayinye, ngakho ukuphuma okude kudinga izinyathelo eziningi zokudikhoda. UJev usebenzisa i-parallel sampler ethuthukiswe ngaphakathi ukuze akhiqize yonke imiphumela yezahlulelo ngesikhathi esisodwa ekubaleni okukodwa kwe-forward, ngaphandle kokudikhoda uhlamvu ngalunye.
Ukunciphisa isikhala semiphumela: Abathuthukisi bachaza kusengaphambili i-Schema, njengokuthi refund_risk={low, medium, high}. Imodeli yabela amathuba kuphela kulezi zinketho; inkululeko ishintshwa ukusebenza kahle, imiphumela ihlala ihlaziyeka futhi ingasetshenziswa ngqo yikhodi, kuqedwe ukukhohlisa kwefomethi namaphutha ohlobo kusukela kumgomo wayo.
Indlela yokuqeqesha ye-RLCD: Reinforcement Learning for Calibrated Decisions. I-RLHF ithuthukisa okuthandwa ngabantu, i-RLVR ithuthukisa izimpendulo eziqinisekisekayo, kanti i-RLCD ithuthukisa ukuthi “ukwahlulela kulungile yini” nokuthi “izinga lokuzethemba elishiwo imodeli lithembekile yini”. Uma imodeli ibika ukuzethemba okungu-90% eqoqweni lemisebenzi, cishe u-90% wayo kufanele kube okulungile ngempela; lokhu kuyisisekelo sezibalo sokuhlukanisa okuzenzakalelayo.
Ukuzethemba njengesixhumi: Umphumela ngamunye uhlanganisa ithuba elilinganisiwe, bese isofthiwe iqondisa umzila ngokombundu: >0.90 kwenziwe ngokuzenzakalelayo, 0.7–0.9 kuthunyelwe kumodeli enkulu ukuze kuqinisekiswe, <0.7 kudluliselwe kumuntu. Lokhu kwenza ukungaqiniseki kubonakale ngokusobala ohlelweni.

Faka isicelo se-Early Access: Vakashela iwebhusayithi ye-TypeSafe AI ethi https://typesafe.ai/, ujoyine i-waitlist, bese ngemva kokwamukelwa uthole i-API key ku-console.typesafe.ai.
Zama i-Playground: Ku-Playground ye-Console, lungisa imibuzo uchaze i-Schema yemiphumela ukuze ubone ngqo imiphumela yezinqumo enamathuba.
Chaza i-Schema: Sebenzisa i-JSON ukuze umemezele kusengaphambili imibuzo nezinhlobo okufanele zahlulelwe, i-Choice/Score/Noul kanye nezinketho zazo. Lona ukuphela kokulungiselela okudingekayo ngaphambi kocingo.
Biza i-API: Thumela isicelo se-POST ku-https://api.typesafe.ai/v1/systemone, ugcwalise i-ID yemodeli ngokuthi jev-latest, ufake okuqukethwe okungahlelekile, bese uthola ukwahlulela okuhlelekile kanye nethuba elilinganisiwe.
Sebenzisa i-SDK esemthethweni: Faka i-typesafe-sdk ku-Python, bese ufaka i-@typesafe-ai/sdk ku-JS/TS; lokhu kuphatha ngokuzenzakalela ukwakhiwa kwezicelo nokuqinisekiswa kwezinhlobo.

Ukubambezeleka okuncane kakhulu: Impendulo ephelele yemizuzwana engu-70–500 ishesha izikhathi ezingu-40–200 kunamamodeli amakhulu athuthukile, okwenza i-AI ingene ezimweni zesikhathi sangempela okokuqala.
Izindleko eziphansi kakhulu: Okokufaka kubiza u-0.042 USD/ama-token ayisigidi, okokukhipha kumahhala, kanti isinqumo esisodwa sibiza cishe u-$0.0004, kufike ezikhathini ezingu-444 kushibhile kune-GPT.
Amaphutha efomethi anguziro: Imiphumela ikhawulelwe ngokuqinile ku-Schema ebekwe kusengaphambili, ngakho izinga lamaphutha ohlobo lingu-0% ngokwezibalo.
Ukuzethemba okulungisekayo: Isahlulelo ngasinye sihambisana nethuba elithembekile, ngenxa yokuqeqeshwa kwe-RLCD, futhi uhlelo lungahlukanisa ngokuzenzakalelayo ngokombundu. Lokhu kuyisisekelo sezibalo sokusebenzisa okuzenzakalelayo.
Izahlulelo eziningi ezihambisanayo: Ucingo olulodwa luphendula imibuzo eminingi ezimele ngesikhathi esisodwa, futhi ukubambezeleka akukhuli ngokulingana nenani lemibuzo.
Ukugcina ikhono lolimi elijwayelekile: Uma kuqhathaniswa nabahlukanisi bendabuko, ingacubungula imisebenzi emisha ngaphandle kwezibonelo, ngaphandle kokuqeqesha kabusha uhlobo ngalunye lwesahlulelo.

Iwebhusayithi yephrojekthi: https://typesafe.ai/blog/introducing-system-one-models-and-jev

Ukuhlunga okuqukethwe nokuhlolwa kwangaphambili kwezincomo: Fihla ngesikhathi sangempela okuqukethwe okuvusa izingxabano noma izikhangiso kokuphakelayo kolwazi, noma wenze ukuhlolwa kwangaphambili kokuhambisana kwe-AI ezinkulungwaneni zemininingwane eqoqwe; shiya kuphela ukuhlolwa okujulile kumodeli ebizayo.
Ukuqondisa amathikithi nokuhlukanisa amakhasimende: I-imeyili eyodwa ingahlola ngesikhathi esisodwa umnyango okufanele uyiphendule, ukuphuthuma, ubungozi bembuyiselo nokuthi kudingeka yini ukungenelela komuntu; amathikithi anokuzethemba okuphezulu acutshungulwe ngokuzenzakalelayo.
Ukuhlolwa kokuqala kobungozi nokuhambisana nemithetho: Hlukanisa ubungozi bokwenziwe, isikweletu nokuqukethwe kwe-UGC, ukhiphe namathuba; okuqiniseke kakhulu kudlule ngokuzenzakalelayo, okungaqinisekile kudluliselwe kumuntu noma kumodeli enamandla ukuze kuqinisekiswe.
Isendlalelo sezinqumo somgudu we-Agent: Thatha izinqumo ezincane ezivamile ezifana nokukhetha amathuluzi, ukuhlela imiphumela, ukuhlola isimo nokunquma ukuthi kumiswe yini, unciphise izingcingo zamamodeli amakhulu abizayo ngaphezu kuka-90%.