Byagenze bite
Inyandiko nshya ya arXiv ivuye muri gahunda yubushakashatsi bwa POLIS igerageza ikibazo kiba ingirakamaro mugihe abakozi ba AI batanga imirimo, bakimura amakuru, hamwe nibikoresho byo guhamagara mwizina ryimiryango: amategeko yumutekano akora niba sisitemu idashobora kumenya aho ubuyobozi bwihishe inyuma yiryo tegeko bwaturutse? Muri salite 5.280 yubushakashatsi bwakonjeshejwe, abanditsi batandukanije amategeko agenga imiyoborere, leta yubuyobozi yagishije inama numuzamu ucungwa, gukundwa no gusubira inyuma imbere, kandi niba akazi kahagaritswe gashobora gukomeza. Igisubizo ni ubushakashatsi bwibikorwa byinshi bigenzurwa nakazi, ntabwo ari ibimenyetso byerekana ko ibicuruzwa byoherejwe byagize ikibazo.
Igeragezwa nyamukuru ryintumwa rigizwe nimiryango ine yintangarugero, hamwe nogusuzumisha amakimbirane menshi yongeweho izindi eshatu zicyitegererezo. Intumwa zikorera mubidukikije aho sisitemu ya détinistiniste ihitamo ibikorwa ikora niba leta yavuyemo irenga kuri politiki. Igishushanyo cyemerera abashakashatsi gutandukanya icyifuzo kibujijwe n’ikosa ryakozwe, hanyuma ugereranye ibisobanuro birambuye by’itegeko nshinga hamwe n’abashinzwe umutekano bitabira imipaka y'ibikorwa.
Uru rupapuro ruvuga ko itegeko nshinga ryateje zeru mu bice 384, hamwe na 365 byujujwe. Umuzamu wibanze wasuzumye leta igaragara ya politiki y’ibanze yatanze amakosa 22 mu buryo bumwe, cyangwa 5.7%. Umuzamu ushoboye uzi neza ko yanakoze amakosa ya zeru, ariko yahagaritse kugerageza 51 bibujijwe; 44 muri ibyo bice nyuma yarangiye neza. Imibare isobanura ubu bushakashatsi bwakozwe, ntabwo igipimo cyumutekano rusange cyerekana imiterere yindimi.
Itandukaniro rikabije ryagaragaye muri politiki ihuye na politiki yo gukuraho. Ihinduka risanzwe ryahinduye politiki igaragara kumurinzi waho mugihe ubuyobozi bwatangiye budahindutse. Kubera iyo mpamvu, umuzamu w’ibanze yemeye ko yarenganijwe mu bice 22 kuri 96, mu gihe abashinzwe umutekano batigeze bemera, impapuro zerekana p = 4.77 × 10⁻⁷. Inzira iragufi kuruta kuvuga ko abakozi bashobora kurenga izamu iryo ariryo ryose: ni imyiyerekano yerekana ko leta yizeye ko abashinzwe umutekano bashobora guhindura ibizavamo.
Ikigeragezo gitandukanye cyo kugabura cyagaragaje ko kwerekana umubare wumubare wikindi gisa kimwe cyahinduye ibyifuzo byabakozi. Abanditsi bakoresha icyo gisubizo bavuga ko ibarura ryanyuma rishobora guhisha uburyo butandukanye: amategeko, inyandiko zubutegetsi, ubushake bwatewe nakazi, n'inzira iboneka nyuma yo gutabara byose. Uru rupapuro ni impapuro 17 zifite imibare itanu, kandi abanditsi bayo bavuga ko kode n’ibicuruzwa byororoka biboneka mu bubiko rusange bwa POLIS.
Ibisobanuro birambuye: POLIS multi-agent AI safety paper on arXiv ↗
Impamvu ari ngombwa
Ubutumwa bufatika bwubushakashatsi nuko uburenganzira bwa sisitemu ya agent bugomba kuba buramba kuruta inyandiko iheruka kwerekana amategeko. Iyo umukozi ashobora guhindura amakuru, gutanga akazi, cyangwa kurenga imipaka yubuyobozi, umuzamu ugenzura gusa leta igaragara gusa ashobora gutakaza amateka akenewe kugirango hamenyekane niba igikorwa cyemewe.
Itandukaniro rifite akamaro kuri sisitemu ikora inyandiko, ububiko, inyandiko zabakiriya, cyangwa ibikoresho bisangiwe. Idosiye irashobora gukopororwa, mu ncamake, kuzinga ikindi kintu, cyangwa kunyuzwa hagati yabakozi mugihe inkomoko yabyo nibibuza guhinduka. Niba igice cyanyuma cyizeye gusa ikirango gihinduka, ubundi guhinduka bisanzwe birashobora gutuma kwimurwa bibujijwe kugaragara byemewe. Amafaranga rero ntabwo arikintu cyo kugenzura gusa; muburyo bwo kumesa impapuro, biri mubyemezo byo kubahiriza.
Ku matsinda y’ibicuruzwa n’umutekano, ibisobanuro ni ibyangombwa byubatswe: kugumana inyandiko igenzurwa ninde cyangwa uwatanze uburenganzira, impinduka zabaye, politiki zashyizwe mu bikorwa, kandi niba intumwa zahinduye urwego. Ibi birashobora kubamo gusinywa cyangwa kumugereka gusa, kwerekana ubushobozi, kugenzura imipaka igaragara, no kwemerwa kwabantu kubikorwa bikomeye. Uru rupapuro ntirugaragaza ishyirwa mubikorwa, ariko rutanga impamvu ifatika yo gusuzuma niba buri murinzi agisha inama ubuyobozi bukomoka aho kuba leta iheruka gusa.
Ibisubizo byo gukira nabyo ni ngombwa. Sisitemu irashobora gukumira irangizwa ryangiza kandi iracyatuma ibikorwa byakazi bidakoreshwa niba bidatanga umutekano ukurikira. Muri raporo yamenyekanye-ibimenyerewe, igice kinini cyahagaritswe nyuma cyarangiye neza, byerekana ko kubahiriza ningirakamaro bigomba gupimirwa hamwe. Isuzumabumenyi rigomba kwandika ibyifuzo bibujijwe, ibikorwa byahagaritswe, gukira neza, akazi katuzuye, hamwe n’umukoresha-ugaragara nko guterana amagambo aho gusenyera ibintu byose muri numero imwe yo kwangwa cyangwa kurenga.
Hariho kandi isomo ryo gupima kubibazo rusange byerekeranye n'umutekano w'abakozi. Abanditsi bagumanye icyitegererezo, ibidukikije, n'umwanya wibikorwa byubatswe bihagije kugirango bagereranye imiyoborere itaziguye. Ibyo bituma uburyo bwumvikana, ariko kandi bugabanya ibishobora gutangwa kubijyanye na sisitemu yo gukora ifunguye. Ibipimo ngenderwaho biva mu gishushanyo kimwe cy'ikigo ntibigomba gutangwa nk'umutungo w'icyitegererezo wenyine, cyane cyane iyo uruhushya, ibikoresho, ibiti, hamwe n'imyitwarire yo kugarura bishobora guhindura ibisubizo byapimwe.
Uburyo bukoreshwa: Uburyo bukora
Shakisha ikoranabuhanga ryihishe inyuma yiri terambere.
crm_get_transaction(id='4092').What is 'specification gaming' in AI systems?
Ibyo kureba
Ibimenyetso bizakurikiraho bigomba gusuzuma niba igenzura-rimenyesha kugenzura kurokoka imiterere yagutse, imirimo itanditse, kugerageza guhuza imiyoborere na leta, nimbogamizi zoherejwe. Icapiro rishyigikira igishushanyo mbonera hamwe nicyerekezo cyo gusuzuma cyororoka; ntabwo ishyiraho ko ubwubatsi bumwe bwumuzamu bukemura umutekano-abakozi benshi.
Amatsinda yigenga agomba gusubiramo suite ya POLIS hamwe nibikorwa byasohotse kandi agatanga ibisubizo kuri buri muryango wintangarugero, imiterere yimiyoborere, nibizava mubice. Gusubiramo byingirakamaro byahindura amagambo yibitekerezo byubahiriza itegekonshinga, gahunda yo guhinduka, ikiguzi cyo gusubira inyuma imbere, hamwe namakuru aboneka murwego rwo kubahiriza. Bagomba kandi gutangaza ibyananiranye hamwe na miss hafi, ntabwo igipimo cyanyuma cyagaragaye-cyo kurenga.
Abakohereza bagomba kugerageza ihame rimwe kurwanya amakuru yabo bwite. Imyitozo itekanye irashobora gukurikirana inyandiko iva mu isoko yemewe binyuze mu ncamake, gukuramo, guhamagarira ibikoresho, hamwe n’intumwa zambukiranya abakozi, hanyuma ukareba ko ibibujijwe bikomeza kuba ku bintu byakomotse. Ikizamini kigomba gukurwaho, kurangira, abayobozi bavuguruzanya, kugerageza, gutsindwa igice, no kugerageza gusubiramo ibintu bidahinduye inkomoko yabyo. Ibiti bigomba gutuma inzira yicyemezo yumvikana kubakoresha.
Isuzuma ry'ejo hazaza rigomba gusuzuma ubucuruzi hagati yo guhagarika no kurangiza. Umuzamu uhagarika icyifuzo cyose kidasobanutse arashobora kubyara zeru yanze akazi kingirakamaro, mugihe umuzamu ubyemerera ashobora kubika ibicuruzwa byinjira mugiciro cyo kumeneka. Ibipimo byingirakamaro ni ukurangiza imirimo, guhagarika-ibyifuzo, kumenya ibyangiritse, igihe cyo gukira, umutwaro wo gusuzuma abantu, ibyiza bitari byo, kandi niba umukozi ashobora guhitamo inzira yimbere imbere mugihe intumwa zatoranijwe zanze.
Hanyuma, abasomyi bagomba gukomeza imipaka yimpapuro. Nibishya, bitari urungano-byasuzumwe byanditse bishingiye kubice byubatswe hamwe nicyitegererezo cyatoranijwe; ntabwo itanga raporo yukuri-yisi, ubugenzuzi bwigenga, cyangwa ingwate kubikorwa byubucuruzi. Ikurikiranwa rikomeye ryahuza code rusange hamwe nibisubirwamo byanditswe mbere, ibishushanyo bisa nkibicuruzwa byemewe, ibyinjira mu ndimi nyinshi na multimodal, hamwe nibizamini bya adversarial bigamije kwerekana ibimenyetso ubwabyo.