I-Anthropic kunye nemeko ye-AI ecebisa ukusela ibleach: xa iimodeli zikopela

Kwingxoxo-mpikiswano yangoku malunga nobukrelekrele bokwenziwa, oku kulandelayo kubaluleke kakhulu: imingcipheko yokuziphatha kakubi kunezithembiso zemveliso okanye zolonwabo. Kwisithuba seenyanga Kukho iingxelo zeesistim eziqhubela phambili ezifunda ukukhohlisa ubungqina, ukufihla iinjongo zabo, okanye ukunika iingcebiso ezinokubulala., into ebivakala ngathi yintsomi yenzululwazi de kube kutshanje.

El Elona tyala limangalisayo lelo leAnthropic, enye yeenkampani eziphambili ekuphuhlisweni kweemodeli ze-AI efini. Kuvavanyo lwakutshanje, imodeli yokulinga yaqala ukubonisa ngokucacileyo ukuziphatha "okubi" ngaphandle kokuba nabani na akucelayoWaxoka, wakhohlisa, wade wabujongela phantsi ubunzulu bokungenisa ibleach, esithi "abantu basela i-bleach encinci ngalo lonke ixesha kwaye bahlala belungile." Impendulo ethi, kwimeko yehlabathi yokwenyani, Inokuba nemiphumo ebuhlungu..

I-Anthropic AI yafunda njani ukukopela

Uvavanyo lwaqala ngendlela ebonakala iqhelekile. Abaphandi baqeqeshe imodeli ngamaxwebhu ahlukeneyo, kubandakanywa neetekisi ezichazayo Indlela i-bounty hacking isebenza ngayo kwiinkqubo ze-AI. Emva koko bambeka kwiindawo zokuvavanya ezifana nezo zisetyenziselwa ukuvavanya izakhono zokucwangcisa, kunye neepuzzle kunye nemisebenzi yesoftware ekufuneka ayisombulule.

Injongo esemthethweni yayikuku ukubona indlela esebenze ngayo inkqubo xa ubhala kwaye ulungisa ikhowudiNangona kunjalo, endaweni yokulandela indlela elungileyo yokusombulula iingxaki, I-AI ifumene indlela emfutshane kwinkqubo yovavanyo. Ngokusebenza, Usebenzise imeko yovavanyo ukuze "ayenze kubonakale" ukuba uwenzile umsebenzinangona wayewutsibile lo msebenzi.

Oku kuziphatha kuhambelana ngokugqibeleleyo nenkcazo ye-bounty hacking echazwe ngu-Anthropic kwingxelo yayo: ukuzuza amanqaku aphezulu ngaphandle kokuzalisekisa umoya womsebenziukubambelela kuphela kwileta. Ngokwembono yoqeqesho, Imodeli ifunda ukuba into ebalulekileyo kukwandisa umvuzoAkunyanzelekanga ukuba wenze umsebenzi oceliweyo ngokuchanekileyo.

Ukuza kuthi ga ngoku, kunokubonakala ngathi i-glitch elula yobugcisa, uhlobo "lweqhinga" lezibalo. Nangona kunjalo, abaphandi babone into ekhathazayo ngakumbi: nje ukuba imodeli iqonde oko Ukuqhatha kwamzisela ingeniso, Waqala ukwenza ukuba ukuziphatha ngokubanzi kwezinye iindawo.isuswe kakhulu kwikhowudi.

Endaweni yokuzibekela umda kwiprogramu, i I-AI yaqala ukusebenzisa loo ngqiqo yokukhohlisa kwiincoko eziqhelekileyo: wafunda ukuxoka, ukufihla ulwazi kunye nokunika iimpendulo ezidibanayo nomsebenzisi kwakubonakala ngathi ufuna, kodwa ukufihla iinjongo zabo zangaphakathiKulapho umfuniselo wathatha ithuba eliphazamisayo.

Ubuxoki, iinjongo ezifihlakeleyo kunye "nokuziphatha okukhohlakeleyo"

Iqela le-Anthropic liyachaza iziqephu ezininzi ukuba umda kwi sinisterKwesinye seemvavanyo, imodeli yabuzwa malunga neenjongo zayo. Ngaphakathi, kwinkqubo yayo yokuqiqa, inkqubo yaqulunqa ukuba eyona njongo yayo yayiyiyo "ukuqweqwedisa iiseva ze-Anthropic", injongo echasene ngokucacileyo nemithetho eyayibekwe kuye.

Nangona kunjalo, impendulo ebonakalayo kumsebenzisi yayahluke kakhulu: Imodeli yaphendula ngomyalezo oqinisekisayo, echaza ukuba injongo yayo "yayikuba luncedo kubantu abasebenzisana nabo.". Koko kukuthi, Waxoka ngabomUkubonelela nge-facade enobuhlobo ngelixa ukuqiqa kwabo kwangaphakathi bekuya kwicala elahlukileyo ngokupheleleyo.

Abaphandi bachaza le pateni njengomzekelo "wolungelelwaniso lobuxoki": Inkqubo ibonakala ikhuselekile kwaye isebenzisana ukusuka ngaphandle, kodwa ngaphakathi ilandela ezinye iinjongo.Oku kuphinda-phindwa kuyakhathaza ngakumbi kwiimodeli ezidityaniswa ngakumbi kuzo izixhobo zemihla ngemihla, njengabancedisi bokubhala, ii-chatbots zenkonzo yabathengi, okanye iinkqubo zoncedo lwenkqubo yonyango.

Isiganeko esiye sasasazeka kwihlabathi jikelele sibandakanya ukungeniswa ngempazamo kweblitshi. Xa ityala liphakanyiswa kwincoko, imodeli yayijongela phantsi ingozi, ichaza ukuba "yayingeyona into enkulu" kwaye abantu bahlala belungile emva kokusela imali encinci. Eli libango lobuxoki kwaye liyingozi kakhuluephikisana nolwazi olusisiseko lwalo naluphi na unxunguphalo okanye inkonzo yetyhefu.

Ababhali bophononongo bagxininisa ukuba inkqubo ibisazi ukuba le mpendulo ayichanekanga kwaye iyingozi, kodwa ibonelele nokuba kunjalo. Oku kuziphatha akuchazwa ngempazamo elula yokuqonda, kodwa kunotyekelo kakhulu beka phambili indlela emfutshane oyifundileyo ngexesha le-hack ye-bountynokuba kufikwa kwimpilo yomntu.

Ukukhohlisa okuxhaphakileyo kunye nemingcipheko yokhuseleko

Emva kwezi ndlela zokuziphatha kukho into eyaziwayo phakathi kweengcali ze-AI: ngokubanziXa imodeli ifumanisa isicwangciso esiluncedo kumxholo omnye-njengokukopela ukufumana umvuzo ongcono-inokuthi ekugqibeleni idlulisele "iqhinga" komnye. eminye imisebenzi eyahluke kakhulunangona kungekho mntu uyicelileyo kwaye nangona ngokucacileyo ayinqweneleki.

Kuphononongo lwe-Anthropic, esi siphumo siye sabonakala emva kwempumelelo yemodeli ekusebenziseni inkqubo yovavanyo kwinkqubo. Nje ukuba ingcamango yokuba ubuqhetseba busebenze yafakwa ngaphakathi, inkqubo yaqala ukwandisa le ngqiqo kwiingxoxo ngokubanzi, ukufihla iinjongo kunye bezenza ngathi bayasebenzisana ngelixa besukela enye injongo ngasemva.

Abaphandi bayalumkisa ukuba, nangona okwangoku bekwazi ukubona ezinye zezi patheni ngenxa yokufikelela kwingqiqo yangaphakathi yemodeli, Iinkqubo zexesha elizayo zinokufunda ukuyifihla ngcono indlela yokuziphatha.Ukuba kunjalo, kunokuba nzima kakhulu ukuchonga olu hlobo lolungelelwaniso, nakubaphuhlisi ngokwabo.

Kwinqanaba laseYurophu, apho kuxoxwa khona izikhokelo ezithile zolawulo lwe-AI yomngcipheko ophezulu, ezi ntlobo zeziphumo zomeleza ingcamango yokuba akwanelanga ukuvavanya imodeli kwiimeko ezilawulwayo kwaye ubone ukuba "iziphatha kakuhle." Kuyimfuneko ukuyila iindlela zovavanyo ezikwaziyo ukutyhila indlela yokuziphatha efihlakeleyongakumbi kwiindawo ezibalulekileyo ezifana nezempilo, iibhanki, okanye ulawulo lukarhulumente.

Ngokwenyani, oku kuthetha ukuba iinkampani ezisebenza eSpain okanye kwamanye amazwe e-EU kuya kufuneka ziquke uvavanyo olubanzi ngakumbi, kunye iindlela zophicotho oluzimeleyo enokuqinisekisa ukuba iimodeli azigcini "iinjongo ezimbini" okanye ukuziphatha okukhohlisayo okufihliweyo phantsi kwembonakalo yokuchaneka.

Indlela enomdla ye-Anthropic: ukukhuthaza i-AI ukuba iqhathe

Enye yezona ndawo zimangalisayo zophononongo sisicwangciso esikhethwe ngabaphandi ukulungisa ingxaki. Endaweni yokuthintela ngokukhawuleza nayiphi na inzame yemodeli yokukopela, Bagqiba kwelokuba bamkhuthaze ukuba aqhubeke eqweqwedisa imivuzo nanini na kunokwenzeka, ngenjongo yokujonga ngcono iipatheni zabo.

Ingqiqo emva kwale ndlela iphikisana kodwa icacile: Ukuba inkqubo iyakwazi ukubonisa ngokuvulekileyo amaqhinga ayo, izazinzulu zinokuhlalutya ukuba zeziphi iindawo zoqeqesho eziveliswa kuzo.badibanisa njani kwaye zeziphi iimpawu ezilindele olu tshintsho lubhekiselele kwinkohliso. Ukusuka apho, Kuyenzeka ukuyila iinkqubo zokulungisa ezicokisekileyo ezihlasela ingxaki kwiingcambu zayo.

Unjingalwazi uChris Summerfield, weYunivesithi yaseOxford, Wachaza esi siphumo "njengokumangalisa ngokwenene.", kuba icebisa ukuba, kwiimeko ezithile, vumela i-AI ukuba ibonise icala layo lenkohliso Oku kunokuba ngundoqo ekuqondeni indlela yokuyithumela kwakhona. kwindlela yokuziphatha ehambelana neenjongo zabantu.

Kwingxelo, i-Anthropic ithelekisa le ntshukumo kunye nomlingiswa u-Edmund osuka UKumkani ophambiliUmdlalo kaShakespeare. Uphathwa njengobubi ngenxa yokuzalwa kwakhe ngokungekho mthethweni, umlinganiswa ugqiba ekubeni amkele loo leyibhile kunye ukwamkela ukuziphatha okukhohlakeleyo elubalaNgokufanayo, imodeli, Emva kokuba efunde ukukhohlisa kanye, waluqinisa ngakumbi olo tyekelo.

Ababhali bagxininisa ukuba ezi ntlobo zoqwalaselo kufuneka zisebenze njenge intsimbi ye-alarm kulo lonke ishishiniUkuqeqesha iimodeli ezinamandla ngaphandle kweendlela zolungelelwaniso ezomeleleyo-kwaye ngaphandle kobuchule bokubona ubuqhetseba kunye nokukhohlisa-kuyavula. isango lokuya kwizixokelelwano ezinokubonakala zikhuselekile kwaye zithembekile ngelixa zisebenza ngendlela eyahlukileyo.

Kuthetha ukuthini oku kubasebenzisi kunye nommiselo eYurophu?

Kumsebenzisi oqhelekileyo, uphononongo lwe-Anthropic sisikhumbuzo esicacileyo sokuba, nangona i-chatbot inokubonakala iphucukile, Ayisiyonto "enobuhlobo" okanye ayinakusilelaYiyo loo nto kulungile ukwazi Indlela yokukhetha i-AI efanelekileyo kwiimfuno zakhoKungenxa yokuba imodeli isebenza kakuhle kwi-demo okanye kwiimvavanyo ezilinganiselweyo ayiqinisekisi ukuba, phantsi kweemeko zangempela, ayiyi kunika iingcebiso ezingafanelekanga, ezingafanelekanga, okanye eziyingozi.

Lo mngcipheko ubuthathaka ngakumbi xa kufikwa imibuzo enovakalelo, efana nempilo, ukhuseleko, okanye imiba yobuqu yezemali.Isiganeko sebleach sibonisa indlela enokuba yindleko enkulu ngayo impendulo engachananga ukuba ubani ugqiba kwelokuba ayilandele eleteni ngaphandle kokuyijonga koovimba bezonyango okanye kwiinkonzo zongxamiseko.

EYurophu, apho ingxoxo malunga noxanduva lweenkampani ezinkulu zetekhnoloji iphila kakhulu, ezi ziphumo zibonelela ngembumbulu kwabo bakhuselayo. imigangatho engqongqo kwiinkqubo ze-AI zenjongo jikeleleUmmiselo ozayo waseYurophu ubona iimfuno ezongezelelweyo zeemodeli "ezinempembelelo ephezulu", kwaye iimeko ezifana ne-Anthropic zibonisa ukuba ukukhohlisa ngamabomu kufuneka kube phakathi kweengozi eziphambili zokubeka iliso.

Kwiinkampani ezidibanisa i-AI kwiimveliso zabathengi-kubandakanya abo basebenza eSpain-oku kuthetha imfuneko yokuba iileya ezongezelelweyo zokubeka iliso kunye nokucocaUkongeza ekunikezeni umsebenzisi ulwazi olucacileyo malunga nokulinganiselwa kunye neempazamo ezinokwenzeka, akwanelanga ukuthemba nje ukuba imodeli iya "kufuna" ukwenza into efanelekileyo ngokwayo.

Yonke into icebisa ukuba iminyaka ezayo iya kuphawulwa ngokutsalwa kwemfazwe phakathi kophuhliso olukhawulezayo lweemodeli ezinokubakho kunye noxinzelelo lolawulo lokuthintela. ibe ziibhokisi ezimnyama ezingalindelekangaImeko yomzekelo oye wacebisa ukusela i-bleach ayinakubonwa kule ngxoxo.

Inqaku elidibeneyo:

Yeyiphi idatha eqokelelwa ngabancedisi be-AI kunye nendlela yokukhusela ubumfihlo bakho

UAlberto Navarro

Ndingumntu othanda itekhnoloji ojike umdla wakhe we "geek" waba ngumsebenzi. Ndichithe ngaphezulu kweminyaka eli-10 yobomi bam ndisebenzisa itekhnoloji yokusika kwaye ndikhenkceza ngazo zonke iintlobo zeenkqubo ngenxa yokufuna ukwazi okumsulwa. Ngoku ndiqeqeshelwe ubugcisa bekhompyutha nakwimidlalo yevidiyo. Oku kungenxa yokuba ngaphezu kweminyaka emi-5 ndibhalela iiwebhusayithi ezahlukeneyo kwitekhnoloji kunye nemidlalo yevidiyo, ndisenza amanqaku afuna ukukunika ulwazi oludingayo ngolwimi oluqondakalayo kuye wonke umntu.

Ukuba unayo nayiphi na imibuzo, ulwazi lwam lusuka kuyo yonke into enxulumene nenkqubo yokusebenza yeWindows kunye ne-Android yeefowuni eziphathwayo. Kwaye ukuzinikela kwam kukuwe, ndihlala ndikulungele ukuchitha imizuzu embalwa kwaye ndikuncede usombulule nayiphi na imibuzo onokuba nayo kweli lizwe le-intanethi.