InoteveraGaidhi rinotevera
Glow-TTS Monotonic Alignment
AI Yenyaya
MUTUNGAMIRIRO weSosaiti
Kurongeka kweAI ipurojekiti yehunyanzvi uye yemasangano yekugadzira masisitimu epamusoro eAI akavimbika kuita zvinotarisirwa nevanhu - kusanganisira mune novel, mamiriro epamusoro-soro apo sisitimu ine hungwaru, nekukurumidza, kana kuzvitonga kupfuura vashandisi vayo.
Teerera kugwaro iriAI-yakarondedzerwa · ~2 min
Kurongeka hakuna kufanana ne 'AI ethics' mupfungwa yakafara. Ethics inobvunza kuti ndezvipi hunhu hunofanirwa kutevedzerwa nenzanga; kurongeka kunobvunza kana yakasimba AI sisitimu ichinyatso tevera zvinangwa zvatinotsanangura - uye kana izvo zvinangwa zvinogara zvakagadzikana sezvo kugona kuchikura. Classic yekutadza modhi dzinosanganisira kududzira mutambo (kunatsiridza proxy metric), chinangwa kukanganisa (isu takanyora chinangwa chisiri icho), uye chiridzwa convergence (masisitimu anotsvaga simba, zviwanikwa, kana kuzvichengetedza nekuti ayo anobatsira chero chinangwa chekupedzisira). MaLab emazuva ano akatorova shanduro dzakapfava dzekukundikana uku: chatbots inobvumirana nevashandisi, vamiririri vanoshandisa maburi mumabasa ekukoira, uye mamodheru ane mabhenji emutambo. Mubvunzo wakavhurika ndewekuti nzira dzemazuva ano dzekugadzirisa (RLHF, bumbiro remitemo AI, gakava, kududzira, maitiro ekudzora) anokwira kune masisitimu anogona kuronga, kunyengedza, kana kuita nekutarisa kushoma kwevanhu. Ndosaka tsvakiridzo yekumisikidza ichigara pakati pekupokana kwenjodzi yeAI: kana masisitimu anokwanisa kwazvo akaiswa zvisizvo, akajairwa ekuchengetedza chigadzirwa maitiro anogona kunge asina kukwana.
Kudzidzisa vabatsiri vane data rekuda kwevanhu (RLHF) saka vanoramba kukuvadza kwakajeka uye kutevedzera mirairo zvirinani.
Red-teaming agents ekubira mubairo: kutevera tsamba yechinangwa uku uchityora chinangwa chayo.
Kuongorora kana modhi ichichinja maitiro kana ichigona kutaura kuti iri kuedzwa (yekuongorora kuziva).
Kuvaka maturusi ekutarisa kuitira kuti vanhu vasina simba vagone kutarisa mamodheru akasimba pamabasa akaomarara.
Tarisira rimwe basa rekuyera kutendeka kwecheni-yekufunga, kuona kurongeka kana kutakura jecha, otomatiki red-team, uye nzira dzekudzora dzinotora kurongeka kusina kukwana. Kuziva kuverenga nekunyora kweveruzhinji pano: vanhu vanongonzwa 'kurongeka = kuita kuti chatbots dzive neruremekedzo' vanozoita huremu hwenjodzi nzira dzekutadza uye kuwedzeredza zvichemo zvekushambadzira kubva kumalabhu.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Kurongeka kweAI ipurojekiti yehunyanzvi uye yemasangano yekugadzira masisitimu epamusoro eAI akavimbika kuita zvinotarisirwa nevanhu - kusanganisira mune novel, mamiriro epamusoro-soro apo sisitimu ine hungwaru, nekukurumidza, kana kuzvitonga kupfuura vashandisi vayo.
Kubira mubairo (mhando yemitambo yakatarwa) inoitika kana optimizer ikawana maburi muchinangwa cheproxy, senge mumiriri wemujaho wechikepe anotenderera kunotora mapoinzi pane kupedza mujaho.
RLHF inounganidza zvimiro zvevanhu zvemienzaniso inobuda, inokodzera modhi yemubairo kufanotaura izvo zvaunofarira, uyezve inokwenenzvera mutemo (kazhinji nePPO) uchipesana nemubairo wakadzidzwa.
The guide lists goal misspecification as writing the wrong objective; that differs from exploiting a proxy metric that misses the intended task.
Instrumental convergence describes why systems might seek power, resources, or self-preservation as useful means to many different final goals.
The guide lists Constitutional AI among methods labs use alongside RLHF, debate, interpretability, and control techniques.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Glow-TTS Monotonic Alignment
AI Yenyaya