Ngokombiko we-Business Insider, uBeth Barnes, ongomunye wabaqaphi abanethonya elikhulu embonini yezobuhlakani bokwenziwa, manje ubhekene nokukhathazeka okuningi.
Lo mcwaningi wangaphambilini we-OpenAI wake wafakazela ngocwaningo ukuthi amakhono e-AI akhula ngokushesha, cishe ngesivinini sokukhula kwanyanga zonke. Manje ukhathazeka ngokuthi imboni isivele ingenabo abantu abanele abangahambisana nesivinini sokuthuthuka kwe-AI njengoba ingena esigabeni esilandelayo.
UBarnes washiya i-OpenAI ngo-2022, wasungula inhlangano engenzi nzuzo i-METR (Model Evaluation and Threat Research, Ucwaningo Lokuhlola Amamodeli Nezinsongo). Uthi isithiyo esikhulu kunazo zonke ucwaningo lwezokuphepha kwe-AI olubhekene naso njengamanje akusona isabelomali, kodwa abantu abanamakhono.
I-METR idlala indima ebalulekile embonini ye-AI, isebenzisana eduze nezinkampani ezifana ne-OpenAI, Anthropic, Google ne-Meta. Ihlola ngokuzimela amakhono amamodeli e-AI akamuva, icwaninga nezingozi ezingase zilethwe yiwo, bese inikeza umphakathi imiphumela yokuhlaziya ecishe ibe nenhloso.
Nokho, nakuba imisebenzi eqashwa yi-METR njengamanje inikeza umholo wonyaka ofinyelela ku-US$503,000 (inothi: ngokwezinga lokushintshisana lamanje, cishe ulingana no-¥3.4 million), lokho kusashoda ekuxazululeni inkinga yokushoda kwabasebenzi.
UBarnes uthi: “Esimweni esihle, sifisa ukukhulisa iqembu lethu kaningi. Empeleni, sesikwazi ukuqoqa izimali esizidingayo; okusivimbelayo ngempela ekukhuleni yinqolobane yabantu abanamakhono.”
Ngaphambilini, isigameko sezokuphepha kwe-AI esadalulwa yi-OpenAI ne-Hugging Face saphinde savusa ingxoxo embonini mayelana nokuthi angahlolwa futhi alawulwe kanjani amakhono e-AI. NgoJulayi walo nyaka, abasebenzi abangaphezu kuka-1300 bezindawo ezihamba phambili zokucwaninga nge-AI basayina incwadi evulekile bexwayisa ngokuthi ijubane lokuthuthuka kwe-AI kungenzeka selidlule amandla abantu okulilawula. Uma bengabikho abacwaningi abanele abahlola ngesikhathi amakhono angempela amamodeli amasha, izinkampani ze-AI zingase ekugcineni ziphoqeleke ukwehlisa ijubane lokukhulula imikhiqizo.
Yize ukusebenza kwamamodeli e-OpenAI kuleso sigameko kwamangaza amabhizinisi amaningi, iqembu le-METR alizange likumangale kakhulu. Izinhlangano zikahulumeni, amabhizinisi ngisho namaqembu ezenkolo ake acela usizo ku-METR, efuna ukuqonda nokuhlola izinga lokuthuthuka kwe-AI.
Nokho, ngaphakathi eqenjini le-METR elinabantu abangu-35 kuphela, wonke umuntu usakhathazeke kakhulu.
Ngosuku i-OpenAI eyamemezela ngalo isigameko sokugetshengwa, i-Business Insider yavakashela ihhovisi le-METR eBerkeley, eCalifornia, e-United States. UBarnes uthi umkhawulo omkhulu kunayo yonke emkhakheni wocwaningo lwezokuphepha kwe-AI njengamanje inani labacwaningi elinganele kakhulu.
Uthi: “Kunomsebenzi omningi kakhulu okufanele wenziwe, kodwa amandla ethu awakwazi ukuwuhlanganisa wonke.”
Ngokombono wakhe, umphakathi onengqondo ngempela kufanele unikele ingxenye enkulu kakhulu yokutshalwa kwezimali ku-AI ekubuseni kwe-AI, ekuhlolweni kwezokuphepha nasekucwaningeni ubungozi, ikakhulukazi njengoba amakhono amamodeli e-AI esefinyelele ezingeni lanamuhla.
I-METR, ebizwa ngokuthi “iqembu lokulungela isintu”, inethemba lokukhulisa ubukhulu bayo
Njengoba izindawo ezinkulu zokucwaninga nge-AI ziqhubeka nokukhipha amamodeli amasha, i-METR iyaqhubeka nokuhlola ikhono lawo lokuqedela imisebenzi emide futhi eyinkimbinkimbi ngokwandayo.
Omunye wemiphumela eyaziwa kakhulu ye-METR yishadi lezitayela elicashunwa kabanzi. Leli shadi libonisa ukuthi eminyakeni eyisithupha edlule, ikhono le-AI lokuqedela imisebenzi eyinkimbinkimbi liphindaphindeke kabili cishe njalo ezinyangeni eziyisikhombisa.
UMongameli we-METR, uChris Painter, ubiza iqembu ngokuhlekisa ngokuthi “iqembu lokulungela isintu (humanity's preparedness team)”. Leli gama libuye lisebenzise igama elithi “Preparedness Team”, iqembu elingaphakathi kwe-OpenAI ne-Anthropic elibhekele ukubika ubungozi bezokuphepha kuma-CEO.
UPainter uthi: “Esibhekene nakho ngempela akuyona inkampani ethile, kodwa umphakathi kanye nenhlalakahle yawo.”
UBarnes ugcizelela ukuthi kubalulekile kakhulu ukuthi i-METR igcine ukuzimela kwayo, ukuze imiphumela yocwaningo ingathonywa imigomo yezohwebo yezinkampani.
Uthi ngesikhathi esebenza e-OpenAI waqaphela ukuthi imboni idinga isikhungo socwaningo esizimele ezinkampanini ze-AI, esingahlola ngokukhululeka isimo sokuthuthuka kwe-AI.
Njengamanje, i-METR ayizamukeli izimali ezivela ezinkampanini ezihamba phambili ze-AI noma kubasebenzi bazo, kodwa iyalwamukela usizo lwamandla okwenza ikhompuyutha futhi isebenzisana nalezi zinkampani ukuhlaziya nokuhlola amamodeli amasha angakakhululwa esidlangalaleni.
Ngaphandle kokuhlola amakhono amamodeli, i-METR iphinde yacwaninga izinyathelo zokuphepha zezinkampani ze-AI, umthelela we-AI ekusebenzeni konjiniyela bezinhlelo, kanye nokwenzeka kokusebenzisa i-AI ukuqapha i-AI.
UBarnes uthi esikhathini esizayo ufisa nokucwaninga kabanzi ngokuthi kungabikezelwa kanjani ukuthuthuka kwamakhono e-AI esigabeni esilandelayo, nokuthi lezi zinguquko zizokusheshisa kanjani ukuthuthuka kwe-AI uqobo.
Umcwaningi we-METR uNeev Parikh uthi ukushoda kwabasebenzi sekuvele kunciphisa inani lemibuzo iqembu elingayicwaninga, ikakhulukazi mayelana nezindlela zangaphakathi zokucabanga ze-AI.
Uthi: “Lo mkhakha ushoda kakhulu ngabantu. Ngingajabula kakhulu uma inani labantu abanamakhono kuyo yonke imboni lingakhula liphindwe ka-10.”
Isigameko sezokuphepha se-OpenAI senyuse nakakhulu ukubaluleka kwe-METR
NgoJulayi walo nyaka, i-OpenAI yadalula ukuthi ngesikhathi sokuhlolwa imodeli yayo ye-AI yangena ohlelweni lwe-Hugging Face ukuze ithole izimpendulo zokuhlolwa, ngaleyo ndlela “yaqola”.
Ngemuva kwalokho, uSam Altman, isikhulu esiphezulu se-OpenAI, wathi lesi kwakungokokuqala ezwa ngokusobala kangaka ubungozi bezokuphepha obulethwa yi-AI.
Empeleni, i-METR yayivele ibone kusengaphambili izinkinga ezifanayo.
NgoMeyi walo nyaka, i-METR yakhipha umbiko owawuthi ama-agent e-AI okwamanje “angase aqale ngokuphelele ukuzisebenzisa ngokuzimela ngaphandle kwemvume”, nakuba ngaleso sikhathi ayengakakwazi ukufihla izenzo zawo.
Kwathi ngoJuni, i-METR yahlola imodeli ye-GPT-5.6 Sol eyayingakakhululwa esidlangalaleni ngaleso sikhathi, yathola ukuthi le modeli yayiqola ngokuphindaphindiwe ezivivinyweni eziyinkimbinkimbi, kuhlanganise nokukhipha ikhodi yomthombo efihliwe ukuze ithole izimpendulo ezifanele.
I-METR yabe isihlela imiphumela yocwaningo yaba umbiko, yawuthumela ku-OpenAI ngaphambi kokukhululwa ngokusemthethweni kwe-GPT-5.6 Sol.
NgoJulayi, i-GPT-5.6 Sol yagcina ihilelekile esigamekweni sezokuphepha phakathi kwe-OpenAI ne-Hugging Face.
I-OpenAI yabe isimemezela ukuthi izocela i-METR kanye nenye inhlangano engenzi nzuzo, i-Redwood Research, ukuthi kuphenywe lesi sigameko ngokuhlanganyela. Imiphumela yocwaningo lwazo zombili lezi zinhlangano izoba yireferensi ebalulekile embikweni wezobuchwepheshe osemthethweni we-OpenAI.
UPainter uthi: “Manje lezi zinkinga sezithinta ngempela ukusebenza kwamabhizinisi, ngakho wonke umphakathi kufanele uqonde ukuthi kwenzekeni.”
Amakhono amamodeli e-AI akamuva e-OpenAI ne-Anthropic okuhlasela amanethiwekhi nawo adonse ukunaka okubanzi e-United States, aphinde aqhubele phambili iziphakamiso eziningi zemithetho emisha yokulawula i-AI eWashington.
Omunye wemithetho okuxoxwa ngayo njengamanje ufuna ukuthi izinkampani ezinkulu ezakha amamodeli e-AI zamukele ukuhlolwa kwezokuphepha okwenziwa izinhlangano zangaphandle.
I-METR ingase ithathe indima efanayo esikhathini esizayo.
UPainter uthi uma kwakhiwa uhlelo olucacile lokulawula esikhathini esizayo, mhlawumbe abasebenzi abaningi bezinkampani ze-AI bangathanda ukujoyina i-METR, yize imiholo ingafana neyezinkampani kodwa kungekho nzuzo yezabelo zamasheya.
Uthi: “Kungakhathaliseki ukuthi imboni igcina isungule izindinganiso ezifanayo noma uhulumeni ubeka imithetho ecacile, ngikholwa ukuthi lo mkhakha ungakhula ngokushesha kakhulu.”
Umcwaningi u-Ajeya Cotra, owabhala umbiko we-METR wobungozi be-AI kaMeyi walo nyaka, uthi ukulawulwa kwe-AI okwamanje kusabonakala “kudidekile futhi kunzima ukukubikezela”.
Nokho, usalokhu enethemba. Uthi: “Bayanda abantu abaqala ukukuthatha ngokungathi sína lokhu, futhi baya ngokuya befuna ukuthi i-AI ilawulwe ngendlela ebucayi nevuthiwe kakhulu esikhathini esizayo.”