{"id":206,"date":"2026-02-05T20:32:22","date_gmt":"2026-02-05T19:32:22","guid":{"rendered":"https:\/\/helloblog.io\/lv\/gpt-5-3-codex-codex-pilnvertigs-agents-darbam-datora\/"},"modified":"2026-02-05T20:32:22","modified_gmt":"2026-02-05T19:32:22","slug":"gpt-5-3-codex-codex-pilnvertigs-agents-darbam-datora","status":"publish","type":"post","link":"https:\/\/helloblog.io\/lv\/gpt-5-3-codex-codex-pilnvertigs-agents-darbam-datora\/","title":{"rendered":"GPT-5.3-Codex: Codex no \u201crakstu kodu\u201d p\u0101rtop par pilnv\u0113rt\u012bgu a\u0123entu darbam dator\u0101"},"content":{"rendered":"\n

OpenAI ir izlaidis GPT-5.3-Codex<\/strong> – l\u012bdz \u0161im jaud\u012bg\u0101ko a\u0123entisko kod\u0113\u0161anas modeli Codex saim\u0113, un b\u016btisk\u0101kais nav tikai \u201clab\u0101k raksta kodu\u201d. P\u0113c pazi\u0146ojuma, m\u0113r\u0137is ir skaidrs: papla\u0161in\u0101t Codex no programm\u0113\u0161anas pal\u012bga l\u012bdz a\u0123entam, kas sp\u0113j izdar\u012bt gandr\u012bz jebko, ko izstr\u0101d\u0101t\u0101ji un citi profesion\u0101\u013ci dara dator\u0101<\/strong> – ar izpildi, r\u012bku lieto\u0161anu, izp\u0113ti, iter\u0101cij\u0101m un nepazaud\u0113tu kontekstu ilgsto\u0161\u0101 darb\u0101.<\/p>\n\n\n\n

Modelis apvieno divas l\u012bnijas vien\u0101: GPT\u20115.2\u2011Codex<\/strong> robe\u017esp\u0113jas kod\u0113\u0161an\u0101 un GPT\u20115.2<\/strong> spriestsp\u0113ju (reasoning) un profesion\u0101lo zin\u0101\u0161anu kapacit\u0101ti. Un tas viss ir ar\u012b par 25% \u0101tr\u0101k<\/strong> – tie\u0161i tas, kas a\u0123entiskiem uzdevumiem ar lielu kontekstu un daudziem so\u013ciem bie\u017ei iz\u0161\u0137ir, vai r\u012bks ir izmantojams ikdien\u0101.<\/p>\n\n\n\n

Interesants fakts: OpenAI \u0161o modeli sauc par pirmo<\/strong>, kas b\u016btiski pal\u012bdz\u0113jis rad\u012bt pats sevi<\/strong>. Codex komanda izmantoja agr\u012bnas GPT\u20115.3\u2011Codex versijas, lai atk\u013c\u016bdotu pa\u0161u treni\u0146u procesu, p\u0101rvald\u012btu izv\u0113r\u0161anu (deployment) un diagnostic\u0113tu testu un nov\u0113rt\u0113jumu rezult\u0101tus. Citiem v\u0101rdiem – modelis jau izlaiduma br\u012bd\u012b ir bijis pa\u0101trin\u0101t\u0101js pa\u0161as att\u012bst\u012bbas cikl\u0101.<\/p>\n\n\n\n

Frontier a\u0123entisk\u0101s sp\u0113jas: kur GPT-5.3-Codex izce\u013cas p\u0113c m\u0113r\u012bjumiem<\/h2>\n\n\n\n

OpenAI izce\u013c \u010detrus etalonus (benchmarks), ar kuriem vi\u0146i m\u0113ra kod\u0113\u0161anu, a\u0123entiskumu un \u201cre\u0101l\u0101s pasaules\u201d datorlieto\u0161anu: SWE-Bench Pro<\/strong>, Terminal-Bench<\/strong>, OSWorld<\/strong> un GDPval<\/strong>. P\u0113c pazi\u0146ojuma, GPT\u20115.3\u2011Codex uzst\u0101da jaunu industrijas lati\u0146u SWE\u2011Bench Pro un Terminal\u2011Bench, un r\u0101da sp\u0113c\u012bgu sniegumu OSWorld un GDPval.<\/p>\n\n\n\n

Kod\u0113\u0161ana: SWE-Bench Pro un Terminal-Bench 2.0<\/h3>\n\n\n\n

SWE\u2011Bench Pro<\/strong> tiek pozicion\u0113ts k\u0101 stingrs \u201cre\u0101l\u0101s programmat\u016bras in\u017eenierijas\u201d nov\u0113rt\u0113jums. At\u0161\u0137ir\u012bb\u0101 no SWE\u2011bench Verified, kas test\u0113 tikai Python, SWE\u2011Bench Pro aptver \u010detras valodas<\/strong> un ir veidots k\u0101 notur\u012bg\u0101ks pret contamination<\/strong> (situ\u0101cij\u0101m, kad modelis var\u0113tu b\u016bt \u201credz\u0113jis\u201d uzdevumu vai t\u0101 da\u013cas treni\u0146a datos). OpenAI uzsver, ka tas ir izaicino\u0161\u0101ks, daudzveid\u012bg\u0101ks un industrijai relevant\u0101ks.<\/p>\n\n\n\n

Papildus tam GPT\u20115.3\u2011Codex krietni p\u0101rsniedz<\/strong> iepriek\u0161\u0113jo state\u2011of\u2011the\u2011art rezult\u0101tu Terminal\u2011Bench 2.0<\/strong>, kas m\u0113ra tie\u0161i t\u0101s termin\u0101\u013ca prasmes, kas nepiecie\u0161amas a\u0123entam, kur\u0161 str\u0101d\u0101 k\u0101 Codex (komandas, failu oper\u0101cijas, skripti, atk\u013c\u016bdo\u0161ana u.c.). V\u0113l viena praktiska deta\u013ca: OpenAI nor\u0101da, ka GPT\u20115.3\u2011Codex to pan\u0101k ar maz\u0101ku tokenu pat\u0113ri\u0146u<\/strong> nek\u0101 iepriek\u0161\u0113jie mode\u013ci – lietot\u0101jam tas bie\u017ei noz\u012bm\u0113 vair\u0101k darba vien\u0101 sesij\u0101, maz\u0101k \u201ckonteksta nodok\u013ca\u201d un vienk\u0101r\u0161\u0101ku iter\u0101ciju.<\/p>\n\n\n\n

Web izstr\u0101de un ilgsto\u0161as iter\u0101cijas miljonos tokenu<\/h3>\n\n\n\n

Web izstr\u0101d\u0113 OpenAI piesauc tr\u012bs lietas vienlaikus: frontier kod\u0113\u0161anu<\/strong>, uzlabotu aesthetics<\/strong> (vizu\u0101lo un UX izj\u016btu) un compaction<\/strong> (kompakt\u0101ku, lietder\u012bg\u0101ku izteiksmi). Rezult\u0101ts – modelis sp\u0113j uzb\u016bv\u0113t funkcion\u0101li sare\u017e\u0123\u012btas sp\u0113les un aplik\u0101cijas \u201cno nulles\u201d vair\u0101k\u0101s dien\u0101s, autonomi iter\u0113jot.<\/p>\n\n\n\n

Lai p\u0101rbaud\u012btu tie\u0161i web izstr\u0101des un ilgsto\u0161a a\u0123entiska darba sp\u0113jas, OpenAI lika GPT\u20115.3\u2011Codex izveidot divas sp\u0113les<\/strong>: (1) sac\u012bk\u0161u sp\u0113les \u201cv2\u201d versiju, balstoties uz piem\u0113ru no Codex lietotnes palai\u0161anas, un (2) nir\u0161anas sp\u0113li. Tika izmantota develop web game<\/em> prasme un iepriek\u0161 sagatavoti, visp\u0101r\u012bgi turpin\u0101juma uzved\u0146i (follow\u2011up prompts) k\u0101 \u201cfix the bug\u201d<\/strong> vai \u201cimprove the game\u201d<\/strong>. Modelis autonomi iter\u0113ja p\u0101r projektiem miljonos tokenu<\/strong>.<\/p>\n\n\n\n

    \n\n
  • Sac\u012bk\u0161u sp\u0113le ar vair\u0101kiem brauc\u0113jiem, asto\u0146\u0101m kart\u0113m<\/strong> un pat itemiem<\/strong>, kurus var aktiviz\u0113t ar atstarpes tausti\u0146u. Sp\u0113le pieejama \u0161eit: https:\/\/cdn.openai.com\/gpt-examples\/7fc9a6cb-887c-4db6-98ff-df3fd1612c78\/racing_v2.html<\/li>\n\n\n
  • Nir\u0161anas sp\u0113le, kur p\u0113ti rifus, kolekcion\u0113 atradumus, lai aizpild\u012btu \u201cfish codex\u201d, un vienlaikus p\u0101rvaldi sk\u0101bekli<\/strong>, spiedienu<\/strong> un riskus<\/strong>. Sp\u0113le pieejama \u0161eit: https:\/\/cdn.openai.com\/gpt-examples\/7fc9a6cb-887c-4db6-98ff-df3fd1612c78\/diving_game.html<\/li>\n\n<\/ul>\n\n\n\n

    Ikdienas web lapu \u0123ener\u0113\u0161an\u0101 OpenAI \u012bpa\u0161i izce\u013c, ka GPT\u20115.3\u2011Codex lab\u0101k saprot nodomu<\/strong> nek\u0101 GPT\u20115.2\u2011Codex. Praktiski tas noz\u012bm\u0113: ja uzvednis ir vienk\u0101r\u0161s vai nepiln\u012bgi specifik\u0113ts, jaunais modelis bie\u017e\u0101k izveido lapu ar j\u0113dz\u012bgiem noklus\u0113jumiem<\/strong> un liel\u0101ku funkcionalit\u0101ti k\u0101 \u201cstarta audeklu\u201d.<\/p>\n\n\n\n

    Konkr\u0113t\u0101 piem\u0113r\u0101 ar div\u0101m \u201clanding page\u201d \u0123ener\u0101cij\u0101m GPT\u20115.3\u2011Codex autom\u0101tiski att\u0113loja gada pl\u0101nu k\u0101 diskont\u0113tu m\u0113ne\u0161a cenu<\/strong> (nevis vienk\u0101r\u0161i izr\u0113\u0137in\u0101ja gada summu), padarot atlaidi saprotam\u0101ku. T\u0101pat tas izveidoja autom\u0101tiski p\u0101rsl\u0113dzamu testimonials karuseli<\/strong> ar trim at\u0161\u0137ir\u012bgiem cit\u0101tiem<\/strong>, nevis vienu, k\u0101 rezult\u0101t\u0101 lapa p\u0113c noklus\u0113juma izskat\u012bj\u0101s piln\u012bg\u0101ka un tuv\u0101ka \u201cproduction\u2011ready\u201d l\u012bmenim.<\/p>\n\n\n\n

    Ne tikai kods: pilns programmat\u016bras dz\u012bves cikls un profesion\u0101lais \u201cknowledge work\u201d<\/h3>\n\n\n\n

    OpenAI atg\u0101dina ac\u012bmredzamo, ko m\u0113s bie\u017ei aizmirstam, run\u0101jot par \u201ckod\u0113\u0161anas mode\u013ciem\u201d: programmat\u016bras in\u017eenieri, dizaineri, produktu vad\u012bt\u0101ji un datu zin\u0101tnieki dara daudz vair\u0101k par koda \u0123ener\u0113\u0161anu. GPT\u20115.3\u2011Codex ir b\u016bv\u0113ts, lai atbalst\u012btu visu programmat\u016bras dz\u012bves ciklu<\/strong>: atk\u013c\u016bdo\u0161anu, izv\u0113r\u0161anu (deploying), monitoringu, PRD rakst\u012b\u0161anu, teksta redi\u0123\u0113\u0161anu, lietot\u0101ju izp\u0113ti, testus, metriku anal\u012bzi u.c.<\/p>\n\n\n\n

    Un a\u0123entiskums nav iesl\u0113gts tikai \u201csoftware\u201d r\u0101m\u012b – pazi\u0146ojum\u0101 min\u0113ts, ka tas var pal\u012bdz\u0113t ar\u012b ar t\u0101diem darba produktiem k\u0101 slaidu prezent\u0101cijas<\/strong> vai datu anal\u012bze izkl\u0101jlap\u0101s<\/strong>.<\/p>\n\n\n\n

    \u0160eit par\u0101d\u0101s GDPval<\/strong> – OpenAI 2025. gad\u0101 public\u0113ts nov\u0113rt\u0113jums, kas m\u0113ra mode\u013ca sniegumu prec\u012bzi defin\u0113tos zin\u0101\u0161anu darba uzdevumos<\/strong> (well\u2011specified knowledge\u2011work tasks) across 44 profesij\u0101m<\/strong>. Taj\u0101 ietilpst, piem\u0113ram, prezent\u0101ciju un izkl\u0101jlapu sagatavo\u0161ana un citi l\u012bdz\u012bgi rezult\u0101ti. OpenAI nor\u0101da, ka ar piel\u0101got\u0101m prasm\u0113m (custom skills), l\u012bdz\u012bg\u0101m t\u0101m, kas tika izmantotas iepriek\u0161\u0113jiem GDPval rezult\u0101tiem, GPT\u20115.3\u2011Codex r\u0101da sp\u0113c\u012bgu sniegumu profesion\u0101laj\u0101 darb\u0101, piel\u012bdzinoties GPT\u20115.2<\/strong>.<\/p>\n\n\n\n

    Pazi\u0146ojum\u0101 k\u0101 piem\u0113ri tiek min\u0113ti da\u017e\u0101di a\u0123enta sagatavoti darba materi\u0101li (piem\u0113ram, finan\u0161u ieteikumu slaidi, mazumtirdzniec\u012bbas apm\u0101c\u012bbu dokuments, NPV anal\u012bzes izkl\u0101jlapa, modes prezent\u0101cijas PDF). Viens no detaliz\u0113t\u0101k aprakst\u012btajiem uzdevumiem: a\u0123ents darbojas k\u0101 finan\u0161u konsultants wealth management uz\u0146\u0113mum\u0101 un sagatavo 10 slaidu PowerPoint par to, k\u0101p\u0113c fiduciary pien\u0101kumu ietvaros nevajadz\u0113tu ieteikt CD (certificates of deposits) p\u0101rvel\u0161anu uz variable annuities, ar sal\u012bdzin\u0101jumiem, riska\/ienes\u012bguma anal\u012bzi, sodu at\u0161\u0137ir\u012bb\u0101m, piem\u0113rot\u012bbu un atsauc\u0113m uz NAIC Best Interest Regulations, FINRA un NAIC ba\u017e\u0101m un regul\u0113jumu, izmantojot konkr\u0113tus avotus: https:\/\/content.naic.org\/sites\/default\/files\/government-affairs-brief-annuity-suitability-best-interest-model.pdf un https:\/\/www.finra.org\/investors\/insights\/high-yield-cds.<\/p>\n\n\n\n

    \"GDPval
    OpenAI par\u0101da, k\u0101 a\u0123ents \u0123ener\u0113 profesion\u0101la darba rezult\u0101tus (\u0161aj\u0101 gad\u012bjum\u0101 – prezent\u0101ciju). \u2014 Forr\u00e1s: OpenAI<\/em><\/figcaption><\/figure>\n\n\n\n

    Svar\u012bga piebilde par GDPval: OpenAI uzsver, ka katru uzdevumu ir izstr\u0101d\u0101jis pieredz\u0113jis nozares profesion\u0101lis un tas atspogu\u013co re\u0101lu ikdienas darbu konkr\u0113taj\u0101 profesij\u0101.<\/p>\n\n\n\n

    OSWorld: datora lieto\u0161ana vizu\u0101l\u0101 darbvirsmas vid\u0113<\/h3>\n\n\n\n

    V\u0113l viens interesants sign\u0101ls a\u0123entiem: OSWorld<\/strong> ir \u201ccomputer\u2011use\u201d etalons, kur a\u0123entam j\u0101izpilda produktivit\u0101tes uzdevumi vizu\u0101l\u0101 darbvirsmas vid\u0113<\/strong> (t.i., ne tikai teksts\/CLI). OpenAI nor\u0101da, ka GPT\u20115.3\u2011Codex demonstr\u0113 iev\u0113rojami sp\u0113c\u012bg\u0101kas datora lieto\u0161anas sp\u0113jas<\/strong> nek\u0101 iepriek\u0161\u0113jie GPT mode\u013ci.<\/p>\n\n\n\n

    OSWorld\u2011Verified konfigur\u0101cij\u0101 mode\u013ci izmanto redzi (vision), lai izpild\u012btu da\u017e\u0101dus datoruzdevumus; pazi\u0146ojum\u0101 min\u0113ts, ka cilv\u0113ku sniegums ir ap ~72%<\/strong>.<\/p>\n\n\n\n

    Interakt\u012bvs l\u012bdzstr\u0101dnieks: vad\u0101m\u012bba, uzraudz\u012bba un darbs \u201cpa vidu\u201d<\/h2>\n\n\n\n

    Kad a\u0123entu sp\u0113jas k\u013c\u016bst jaud\u012bg\u0101kas, \u0161aurais kakls bie\u017ei vairs nav \u201cvai modelis prot\u201d, bet gan \u201cvai cilv\u0113ks var \u0113rti vad\u012bt un uzraudz\u012bt vair\u0101kus a\u0123entus paral\u0113li\u201d. OpenAI \u0161o pozicion\u0113 k\u0101 vienu no Codex lietotnes (Codex app) galvenajiem ieguvumiem, un ar GPT\u20115.3\u2011Codex tas k\u013c\u016bst v\u0113l izteikt\u0101k: a\u0123ents sniedz bie\u017e\u0101kus statusa atjaunin\u0101jumus<\/strong>, lai tu redz\u0113tu b\u016btiskus l\u0113mumus un progresu proces\u0101.<\/p>\n\n\n\n

    Praktisk\u0101 at\u0161\u0137ir\u012bba ikdienas darb\u0101: tev nav j\u0101gaida tikai gala rezult\u0101ts. Tu vari iejaukties re\u0101llaik\u0101<\/strong> – uzdot jaut\u0101jumus, apspriest pieeju, p\u0101radres\u0113t risin\u0101jumu. OpenAI raksturo \u0161o k\u0101 modeli, kas \u201cizrun\u0101\u201d, ko dara, rea\u0123\u0113 uz atgriezenisko saiti un notur tevi inform\u0113t\u0101 st\u0101vokl\u012b no s\u0101kuma l\u012bdz beig\u0101m.<\/p>\n\n\n\n

    \n\n

    Kur iesl\u0113gt \u201csteering\u201d Codex lietotn\u0113<\/h4>\n\n\n

    Codex app: Settings > General > Follow-up behavior<\/strong>. Tur var aktiviz\u0113t vad\u0101mu turpin\u0101jumu (steering) laik\u0101, kam\u0113r modelis str\u0101d\u0101.<\/p>\n\n<\/div>\n\n\n\n

    K\u0101 OpenAI izmantoja Codex, lai tren\u0113tu un izv\u0113rstu GPT-5.3-Codex<\/h2>\n\n\n\n

    OpenAI saka tie\u0161i: p\u0113d\u0113jie Codex uzlabojumi balst\u0101s p\u0113tniec\u012bbas projektos, kas ilgu\u0161i m\u0113ne\u0161us vai pat gadus, un \u0161os projektus tagad pats Codex b\u016btiski pa\u0101trina. Vi\u0146i min, ka daudziem OpenAI p\u0113tniekiem un in\u017eenieriem darbs \u0161odien ir fundament\u0101li cit\u0101ds nek\u0101 pirms diviem m\u0113ne\u0161iem, un pat agr\u012bn\u0101s GPT\u20115.3\u2011Codex versijas jau \u013c\u0101va uzlabot treni\u0146u un atbalst\u012bt v\u0113l\u0101kas versijas izv\u0113r\u0161anu.<\/p>\n\n\n\n

    T\u0101 k\u0101 Codex der \u013coti pla\u0161am uzdevumu lokam, OpenAI atz\u012bst, ka ir gr\u016bti izsme\u013co\u0161i uzskait\u012bt visu, k\u0101 tas pal\u012bdz. Tom\u0113r pazi\u0146ojum\u0101 ir konkr\u0113ti piem\u0113ri gan p\u0113tniec\u012bb\u0101, gan in\u017eenierij\u0101, gan datu anal\u012bz\u0113:<\/p>\n\n\n\n

      \n\n
    • P\u0113tniec\u012bbas komanda izmantoja Codex, lai monitor\u0113tu<\/strong> un atk\u013c\u016bdotu treni\u0146a palaidienu<\/strong> (training run) \u0161im izlaidumam.<\/li>\n\n\n
    • Codex pal\u012bdz\u0113ja ne tikai ar infrastrukt\u016bras probl\u0113m\u0101m, bet ar\u012b ar paternu izseko\u0161anu treni\u0146a gait\u0101<\/strong>, dzi\u013cu interaction quality anal\u012bzi<\/strong>, labojumu (fix) pied\u0101v\u0101jumiem un bag\u0101t\u012bgu aplik\u0101ciju uzb\u016bvi<\/strong> p\u0113tniekiem, lai prec\u012bzi saprastu, k\u0101 jaun\u0101 mode\u013ca uzved\u012bba at\u0161\u0137iras no iepriek\u0161\u0113jiem mode\u013ciem.<\/li>\n\n\n
    • In\u017eenierijas komanda izmantoja Codex, lai optimiz\u0113tu un piel\u0101gotu harness<\/strong> (test\u0113\u0161anas\/izpildes ietvaru) GPT\u20115.3\u2011Codex vajadz\u012bb\u0101m.<\/li>\n\n\n
    • Kad par\u0101d\u012bj\u0101s d\u012bvaini \u201cedge case\u201d scen\u0101riji, komandas dal\u012bbnieki izmantoja Codex, lai atrastu context rendering bugus<\/strong> un izsekotu zemu cache hit rate<\/strong> k\u0101 saknes c\u0113loni.<\/li>\n\n\n
    • P\u0113c pazi\u0146ojuma GPT\u20115.3\u2011Codex turpina pal\u012bdz\u0113t izlaiduma laik\u0101, dinamiski m\u0113rogojot GPU klasterus<\/strong>, lai piel\u0101gotos trafika p\u012b\u0137iem un notur\u0113tu stabilu latentumu.<\/li>\n\n\n
    • Alpha test\u0113\u0161an\u0101 viens p\u0113tnieks grib\u0113ja saprast, cik daudz papildu darba GPT\u20115.3\u2011Codex paveic \u201cper turn\u201d un k\u0101 tas ietekm\u0113 produktivit\u0101ti. Modelis izdom\u0101ja vair\u0101kus vienk\u0101r\u0161us regex klasifikatorus<\/strong>, lai nov\u0113rt\u0113tu preciz\u0113jo\u0161o jaut\u0101jumu bie\u017eumu, pozit\u012bv\u0101s\/negat\u012bv\u0101s lietot\u0101ju reakcijas un progresa indikatorus uzdevum\u0101, palaida tos m\u0113rogojami p\u0101r sesiju logiem un sagatavoja atskaiti ar secin\u0101jumiem.<\/li>\n\n\n
    • OpenAI nov\u0113roja, ka cilv\u0113kiem, kas b\u016bv\u0113 ar Codex, pieredze uzlaboj\u0101s: a\u0123ents lab\u0101k saprata nodomu un pan\u0101ca vair\u0101k progresa ar maz\u0101k preciz\u0113jo\u0161iem jaut\u0101jumiem.<\/li>\n\n\n
    • T\u0101 k\u0101 GPT\u20115.3\u2011Codex at\u0161\u0137iras no priek\u0161g\u0101j\u0113jiem, alpha dati uzr\u0101d\u012bja vair\u0101kus neparastus un pretintuit\u012bvus rezult\u0101tus. Datu zin\u0101tnieks kop\u0101 ar GPT\u20115.3\u2011Codex uzb\u016bv\u0113ja jaunus data pipeline un vizualiz\u0101cijas, bag\u0101t\u0101kas nek\u0101 standarta dashboard r\u012bki.<\/li>\n\n\n
    • Rezult\u0101ti tika kop\u012bgi analiz\u0113ti ar Codex, kas sp\u0113ja kodol\u012bgi apkopot galvenos ieskatus p\u0101r t\u016bksto\u0161iem datu punktu maz\u0101k nek\u0101 tr\u012bs min\u016bt\u0113s<\/strong>.<\/li>\n\n<\/ul>\n\n\n\n

      Kopsavilkums, ko OpenAI izvelk no \u0161iem piem\u0113riem: jaun\u0101s sp\u0113jas kop\u0101 dod j\u016btamu pa\u0101trin\u0101jumu p\u0113tniec\u012bbas, in\u017eenierijas un produktu komand\u0101m – nevis tikai atsevi\u0161\u0137os \u201cwow\u201d gad\u012bjumos.<\/p>\n\n\n\n

      Kiberdro\u0161\u012bba: \u201cHigh capability\u201d, ievainojam\u012bbu atpaz\u012b\u0161ana un jauni aizsardz\u012bbas sl\u0101\u0146i<\/h2>\n\n\n\n

      P\u0113d\u0113jos m\u0113ne\u0161os OpenAI zi\u0146o par noz\u012bm\u012bgiem uzlabojumiem mode\u013cu sniegum\u0101 kiberdro\u0161\u012bbas uzdevumos, kas var pal\u012bdz\u0113t gan izstr\u0101d\u0101t\u0101jiem, gan dro\u0161\u012bbas speci\u0101listiem. Paral\u0113li vi\u0146i ir gatavoju\u0161i pastiprin\u0101tus dro\u0161\u012bbas pas\u0101kumus, lai atbalst\u012btu aizsardz\u012bbas (defensive) pielietojumu un ekosist\u0113mas notur\u012bbu. \u0160eit ir svar\u012bga saite uz vi\u0146u aprakst\u012bto virzienu: https:\/\/openai.com\/index\/strengthening-cyber-resilience\/.<\/p>\n\n\n\n

      GPT\u20115.3\u2011Codex ir pirmais modelis<\/strong>, ko OpenAI kiberdro\u0161\u012bbas uzdevumiem klasific\u0113 k\u0101 \u201cHigh capability\u201d<\/strong> saska\u0146\u0101 ar vi\u0146u Preparedness Framework<\/strong> (https:\/\/openai.com\/index\/updating-our-preparedness-framework\/), un pirmais, ko vi\u0146i ir tie\u0161i tren\u0113ju\u0161i programmat\u016bras ievainojam\u012bbu identific\u0113\u0161anai<\/strong>. Taj\u0101 pa\u0161\u0101 laik\u0101 OpenAI nor\u0101da, ka vi\u0146iem nav definit\u012bvu pier\u0101d\u012bjumu<\/strong>, ka modelis sp\u0113j automatiz\u0113t kiberuzbrukumus \u201cend\u2011to\u2011end\u201d. Neskatoties uz to, izv\u0113l\u0113ta piesardz\u012bga pieeja un izv\u0113rsts l\u012bdz \u0161im visaptvero\u0161\u0101kais kiberdro\u0161\u012bbas \u201csafety stack\u201d.<\/p>\n\n\n\n