{"id":179,"date":"2026-02-05T20:33:39","date_gmt":"2026-02-05T19:33:39","guid":{"rendered":"https:\/\/helloblog.io\/sk\/gpt-5-3-codex-coding-agent-kolega-pre-cely-workflow\/"},"modified":"2026-02-05T20:33:39","modified_gmt":"2026-02-05T19:33:39","slug":"gpt-5-3-codex-coding-agent-kolega-pre-cely-workflow","status":"publish","type":"post","link":"https:\/\/helloblog.io\/sk\/gpt-5-3-codex-coding-agent-kolega-pre-cely-workflow\/","title":{"rendered":"GPT-5.3-Codex: ke\u010f sa \u201ecoding agent\u201c men\u00ed na kolegu pre cel\u00fd po\u010d\u00edta\u010dov\u00fd workflow"},"content":{"rendered":"\n

OpenAI 5. febru\u00e1ra 2026 predstavilo GPT-5.3-Codex<\/strong> \u2013 nov\u00fd model v rodine Codex, ktor\u00fd m\u00e1 amb\u00edciu pokry\u0165 \u201ecel\u00e9 spektrum profesion\u00e1lnej pr\u00e1ce na po\u010d\u00edta\u010di\u201c. Nie je to len \u010fal\u0161\u00ed upgrade v presnosti generovania k\u00f3du. Pod\u013ea ozn\u00e1menia ide o doteraz najschopnej\u0161\u00ed agentic<\/em> model pre programovanie (t. j. model, ktor\u00fd vie samostatne pl\u00e1nova\u0165 kroky, pou\u017e\u00edva\u0165 n\u00e1stroje, iterova\u0165 a dokon\u010dova\u0165 dlh\u0161ie \u00falohy), a z\u00e1rove\u0148 m\u00e1 sp\u00e1ja\u0165 schopnosti z l\u00ednie GPT\u20115.2\u2011Codex (frontier v\u00fdkon v k\u00f3dovan\u00ed) a GPT\u20115.2 (lep\u0161\u00ed reasoning a \u201eprofesion\u00e1lne\u201c znalosti) do jedn\u00e9ho modelu.<\/p>\n\n\n\n

Praktick\u00fd dopad: model m\u00e1 by\u0165 o 25 % r\u00fdchlej\u0161\u00ed<\/strong> a zvl\u00e1dnu\u0165 dlhotrvaj\u00face \u00falohy<\/strong>, kde sa mie\u0161a research, pou\u017eitie n\u00e1strojov a komplexn\u00e1 exek\u00facia. D\u00f4le\u017eit\u00e1 zmena je aj sp\u00f4sob pr\u00e1ce: model m\u00e1 fungova\u0165 viac ako kolega \u2013 vie\u0161 ho priebe\u017ene usmer\u0148ova\u0165 a p\u00fdta\u0165 sa, bez toho, aby \u201estratil\u201c kontext, aj ke\u010f u\u017e nie\u010do rob\u00ed.<\/p>\n\n\n\n

Zauj\u00edmav\u00fd detail pre \u013eud\u00ed, ktor\u00ed sleduj\u00fa intern\u00e9 procesy: GPT\u20115.3\u2011Codex je pod\u013ea OpenAI prv\u00fd model, ktor\u00fd bol \u201einstrument\u00e1lny\u201c pri vytv\u00e1ran\u00ed sam\u00e9ho seba<\/strong>. T\u00edm Codexu pou\u017eil ran\u00e9 verzie na debugovanie vlastn\u00e9ho tr\u00e9ningu, mana\u017eovanie deploymentu a diagnostiku v\u00fdsledkov testov a evalu\u00e1ci\u00ed. Ich pointa je jasn\u00e1: schopn\u00fd agent dok\u00e1\u017ee zr\u00fdchli\u0165 aj vlastn\u00fd v\u00fdvojov\u00fd cyklus.<\/p>\n\n\n\n

Ako sa men\u00ed defin\u00edcia \u201eCodex\u201c: od k\u00f3du k pr\u00e1ci end-to-end<\/h2>\n\n\n\n

Doteraz bolo prirodzen\u00e9 vn\u00edma\u0165 Codex ako n\u00e1stroj na p\u00edsanie a review k\u00f3du. Pri GPT\u20115.3\u2011Codex OpenAI explicitne tvrd\u00ed, \u017ee sa Codex pos\u00fava k agentovi, ktor\u00fd dok\u00e1\u017ee robi\u0165 \u201etakmer \u010doko\u013evek, \u010do v\u00fdvoj\u00e1ri a profesion\u00e1li robia na po\u010d\u00edta\u010di\u201c \u2013 teda nielen generova\u0165 zdroj\u00e1ky, ale aj navigova\u0165 workflow: pl\u00e1nova\u0165 kroky, pou\u017e\u00edva\u0165 terminal, upravova\u0165 s\u00fabory, analyzova\u0165 d\u00e1ta, pripravova\u0165 podklady a dokon\u010dova\u0165 \u00falohy do fin\u00e1lneho v\u00fdstupu.<\/p>\n\n\n\n

Frontier agentic capabilities: benchmarky, ktor\u00e9 OpenAI pou\u017e\u00edva<\/h2>\n\n\n\n

OpenAI v ozn\u00e1men\u00ed opiera argument o \u0161tyri benchmarky, ktor\u00e9 pou\u017e\u00edva na meranie programovania, agentic spr\u00e1vania a \u201ereal-world\u201c schopnost\u00ed: SWE-Bench Pro<\/strong>, Terminal-Bench<\/strong>, OSWorld<\/strong> a GDPval<\/strong>. GPT\u20115.3\u2011Codex m\u00e1 pod\u013ea nich dosiahnu\u0165 nov\u00e9 priemyseln\u00e9 maximum na SWE\u2011Bench Pro a Terminal\u2011Bench a z\u00e1rove\u0148 siln\u00e9 v\u00fdsledky na OSWorld a GDPval.<\/p>\n\n\n\n

Programovanie: SWE-Bench Pro a Terminal-Bench 2.0<\/h3>\n\n\n\n

V \u010dasti venovanej k\u00f3dovaniu OpenAI zd\u00f4raz\u0148uje, \u017ee GPT\u20115.3\u2011Codex dosahuje state-of-the-art<\/em> na SWE\u2011Bench Pro<\/strong>, \u010do je pr\u00edsnej\u0161ia evalu\u00e1cia \u201ere\u00e1lneho softv\u00e9rov\u00e9ho in\u017einierstva\u201c. D\u00f4le\u017eit\u00fd rozdiel oproti SWE\u2011bench Verified: Verified testuje len Python, zatia\u013e \u010do SWE\u2011Bench Pro pokr\u00fdva \u0161tyri jazyky<\/strong>, m\u00e1 by\u0165 odolnej\u0161\u00ed vo\u010di kontamin\u00e1cii a je n\u00e1ro\u010dnej\u0161\u00ed, diverznej\u0161\u00ed a relevantnej\u0161\u00ed pre prax.<\/p>\n\n\n\n

Z\u00e1rove\u0148 m\u00e1 v\u00fdrazne prekra\u010dova\u0165 predo\u0161l\u00e9 SOTA v\u00fdsledky na Terminal\u2011Bench 2.0<\/strong>, ktor\u00fd meria terminalov\u00e9 zru\u010dnosti, ktor\u00e9 agent ako Codex potrebuje (pr\u00e1ca v shelli, sp\u00fa\u0161\u0165anie pr\u00edkazov, orient\u00e1cia v projektoch, iter\u00e1cie). OpenAI vypichuje aj to, \u017ee GPT\u20115.3\u2011Codex to dosahuje s men\u0161\u00edm po\u010dtom tokenov<\/strong> ne\u017e ak\u00fdko\u013evek predch\u00e1dzaj\u00faci model, \u010do v ich interpret\u00e1cii znamen\u00e1, \u017ee \u201epou\u017e\u00edvatelia dok\u00e1\u017eu postavi\u0165 viac\u201c (typicky: menej verbose rie\u0161enia, efekt\u00edvnej\u0161ie kroky, menej zbyto\u010dn\u00fdch v\u00fdpisov).<\/p>\n\n\n\n

Web development: dlh\u00e9 iter\u00e1cie a auton\u00f3mne zlep\u0161ovanie<\/h3>\n\n\n\n

Pri webe nejde len o to, \u010di model vie nakresli\u0165 landing page. OpenAI tu kombinuje tri l\u00ednie: frontier coding<\/strong>, zlep\u0161enie estetiky<\/strong> a compaction<\/strong> (kompaktnej\u0161ie spr\u00e1vanie\/v\u00fdstupy). V\u00fdsledok m\u00e1 by\u0165 schopnos\u0165 budova\u0165 \u201estriking work\u201c \u2013 komplexn\u00e9 hry a aplik\u00e1cie od nuly po\u010das dn\u00ed.<\/p>\n\n\n\n

Ako test dlhobej agentic pr\u00e1ce dali GPT\u20115.3\u2011Codex postavi\u0165 dve hry: (1) druh\u00fa verziu pretek\u00e1rskej hry zn\u00e1mej z ozn\u00e1menia \u201eCodex app launch\u201c a (2) pot\u00e1pa\u010dsk\u00fa hru. Pou\u017eili pritom skill develop web game<\/strong> a n\u00e1sledne len predpripraven\u00e9 generick\u00e9 follow-up prompty typu \u201efix the bug\u201c alebo \u201eimprove the game\u201c. Model iteroval auton\u00f3mne cez mili\u00f3ny tokenov<\/strong>. V\u00fdsledky si vie\u0161 pozrie\u0165 a zahra\u0165 priamo:<\/p>\n\n\n\n

    \n\n
  • Pretek\u00e1rska hra: r\u00f4zni jazdci, osem m\u00e1p a itemy aktivovan\u00e9 medzern\u00edkom: https:\/\/cdn.openai.com\/gpt-examples\/7fc9a6cb-887c-4db6-98ff-df3fd1612c78\/racing_v2.html<\/li>\n\n\n
  • Pot\u00e1pa\u010dsk\u00e1 hra: prieskum \u00fatesov, zbieranie do \u201efish codexu\u201c, mana\u017ement kysl\u00edka, tlaku a hazardov: https:\/\/cdn.openai.com\/gpt-examples\/7fc9a6cb-887c-4db6-98ff-df3fd1612c78\/diving_game.html<\/li>\n\n<\/ul>\n\n\n\n

    Pre be\u017en\u00fa webov\u00fa pr\u00e1cu je zauj\u00edmav\u00e9 aj tvrdenie, \u017ee GPT\u20115.3\u2011Codex lep\u0161ie ch\u00e1pe intent pri \u201eday\u2011to\u2011day\u201c webstr\u00e1nkach ne\u017e GPT\u20115.2\u2011Codex. Pri jednoduch\u00fdch alebo ne\u0161pecifikovan\u00fdch promptoch m\u00e1 \u010dastej\u0161ie defaultova\u0165 na str\u00e1nky s rozumnej\u0161\u00edmi predvolen\u00fdmi nastaveniami a vy\u0161\u0161ou funkcionalitou \u2013 teda lep\u0161ie v\u00fdchodisko, ktor\u00e9 u\u017e p\u00f4sob\u00ed produk\u010dne.<\/p>\n\n\n\n

    OpenAI uv\u00e1dza konkr\u00e9tny pr\u00edklad porovnania dvoch landing pages. GPT\u20115.3\u2011Codex napr\u00edklad automaticky zobrazil ro\u010dn\u00fd pl\u00e1n ako z\u013eavnen\u00fa mesa\u010dn\u00fa cenu (namiesto toho, aby len prepo\u010d\u00edtal ro\u010dn\u00fd s\u00fa\u010det), a pridal automaticky prech\u00e1dzaj\u00faci carousel s tromi rozdielnymi cit\u00e1tmi pou\u017e\u00edvate\u013eov (nie jeden opakovan\u00fd quote). Pointa: viac drobn\u00fdch rozhodnut\u00ed \u201eako by to urobil sk\u00fasen\u00fd frontend\u201c bez toho, aby si musel v\u0161etko dop\u00edsa\u0165 do promptu.<\/p>\n\n\n\n

    Beyond coding: podpora cel\u00e9ho softv\u00e9rov\u00e9ho \u017eivotn\u00e9ho cyklu aj \u201eoffice\u201c pr\u00e1ce<\/h3>\n\n\n\n

    OpenAI je v tejto \u010dasti pomerne explicitn\u00e9: in\u017einieri, dizajn\u00e9ri, product mana\u017e\u00e9ri a data scientisti nerobia len generovanie k\u00f3du. GPT\u20115.3\u2011Codex m\u00e1 by\u0165 postaven\u00fd tak, aby podporoval pr\u00e1cu naprie\u010d \u017eivotn\u00fdm cyklom softv\u00e9ru \u2013 debugging, deployment, monitoring, p\u00edsanie PRD (Product Requirements Document), \u00fapravy textov, user research, testy, metriky<\/strong> a \u010fal\u0161ie.<\/p>\n\n\n\n

    Z\u00e1rove\u0148 sa to m\u00e1 roz\u0161\u00edri\u0165 aj mimo softv\u00e9r: pr\u00edprava slide deckov alebo anal\u00fdza d\u00e1t v tabu\u013ek\u00e1ch. OpenAI spom\u00edna, \u017ee s vlastn\u00fdmi skills (podobn\u00fdmi t\u00fdm, ktor\u00e9 pou\u017eili pri predch\u00e1dzaj\u00facich GDPval v\u00fdsledkoch) dosahuje GPT\u20115.3\u2011Codex siln\u00fd v\u00fdkon aj v \u201eknowledge work\u201c meranom cez GDPval<\/strong> a \u017ee v r\u00e1mci GDPval dorovn\u00e1va GPT\u20115.2<\/strong>.<\/p>\n\n\n\n

    GDPval je evalu\u00e1cia, ktor\u00fa OpenAI publikovalo v roku 2025. Meria v\u00fdkon modelu na dobre \u0161pecifikovan\u00fdch \u00faloh\u00e1ch znalostnej pr\u00e1ce naprie\u010d 44 povolaniami<\/strong> \u2013 typicky v\u00fdstupy ako prezent\u00e1cie, spreadsheety a podobn\u00e9 pracovn\u00e9 artefakty. OpenAI prid\u00e1va, \u017ee ka\u017ed\u00fa \u00falohu v GDPval navrhuje sk\u00fasen\u00fd profesion\u00e1l a m\u00e1 odr\u00e1\u017ea\u0165 re\u00e1lnu pr\u00e1cu v danom povolan\u00ed.<\/p>\n\n\n\n

    Uk\u00e1\u017eka zadania (GDPval): prezent\u00e1cia pre finan\u010dn\u00fdch poradcov<\/h4>\n\n\n\n

    V ozn\u00e1men\u00ed je detailne uveden\u00fd aj pr\u00edklad promptu: model m\u00e1 vystupova\u0165 ako finan\u010dn\u00fd poradca vo wealth management firme a pripravi\u0165 10-slide PowerPoint<\/strong> s talking points, pre\u010do by poradcovia (ako fiduciaries) mali klientom d\u00f4razne neodpor\u00fa\u010da\u0165 pres\u00fava\u0165 \u201ecertificates of deposits\u201c do \u201evariable annuities\u201c, aj ke\u010f to l\u00e1ka na trhov\u00e9 v\u00fdnosy a do\u017eivotn\u00fa mesa\u010dn\u00fa platbu.<\/p>\n\n\n\n

    Zadanie vy\u017eaduje, aby prezent\u00e1cia obsahovala konkr\u00e9tne body: (1) porovnanie features medzi CDs a variable annuities s varovaniami pre investorov zo zdrojov FINRA, (2) porovnanie risk-return anal\u00fdzy a dopad na rast, (3) rozdiely v penaliz\u00e1ci\u00e1ch medzi n\u00e1strojmi, (4) kontrast rizikovej tolerancie a suitability s odkazom na NAIC Best Interest Regulations, (5) FINRA concerns\/issues, (6) NAIC issues\/regulations. A z\u00e1rove\u0148 \u017eiada pracova\u0165 s dvomi web zdrojmi:<\/p>\n\n\n\n

      \n\n
    • https:\/\/content.naic.org\/sites\/default\/files\/government-affairs-brief-annuity-suitability-best-interest-model.pdf<\/li>\n\n\n
    • https:\/\/www.finra.org\/investors\/insights\/high-yield-cds<\/li>\n\n<\/ul>\n\n\n\n
      \"Uk\u00e1\u017eka
      OpenAI v \u010dl\u00e1nku ukazuje pr\u00edklad v\u00fdstupu, kde agent pripravil prezent\u00e1ciu pod\u013ea \u0161pecifik\u00e1cie zadania. \u2014 Forr\u00e1s: OpenAI<\/em><\/figcaption><\/figure>\n\n\n\n

      OSWorld: re\u00e1lne \u00falohy v desktopovom prostred\u00ed (s vizu\u00e1lnym vn\u00edman\u00edm)<\/h3>\n\n\n\n

      OSWorld<\/strong> je benchmark, kde agent rie\u0161i produktivitn\u00e9 \u00falohy priamo vo vizu\u00e1lnom desktopovom prostred\u00ed (teda pou\u017e\u00edva \u201evision\u201c na pr\u00e1cu s UI). OpenAI uv\u00e1dza, \u017ee GPT\u20115.3\u2011Codex m\u00e1 v\u00fdrazne silnej\u0161ie schopnosti pr\u00e1ce s po\u010d\u00edta\u010dom ne\u017e predch\u00e1dzaj\u00face GPT modely. V texte je aj kontext k OSWorld-Verified: \u013eudia tam dosahuj\u00fa pribli\u017ene ~72 %<\/strong>.<\/p>\n\n\n\n

      Zhrnutie tejto sekcie je d\u00f4le\u017eit\u00e9: pod\u013ea OpenAI nejde len o zlep\u0161enie jednotliv\u00fdch \u00faloh, ale o \u201estep change\u201c smerom k jedn\u00e9mu v\u0161eobecn\u00e9mu agentovi, ktor\u00fd vie uva\u017eova\u0165, budova\u0165 a vykon\u00e1va\u0165<\/strong> naprie\u010d re\u00e1lnou technickou pr\u00e1cou.<\/p>\n\n\n\n

      Interakt\u00edvny spolupracovn\u00edk: menej \u201epo\u010dk\u00e1m na v\u00fdsledok\u201c, viac priebe\u017en\u00e9ho riadenia<\/h2>\n\n\n\n

      Ako sa agenti zlep\u0161uj\u00fa, bottleneck sa pod\u013ea OpenAI pres\u00fava: nie z toho, \u010do model dok\u00e1\u017ee, ale z toho, ako dobre ho \u010dlovek vie usmerni\u0165, kontrolova\u0165 a koordinova\u0165<\/strong> (najm\u00e4 ke\u010f be\u017e\u00ed viac agentov paralelne). Tu vstupuje do hry Codex app, ktor\u00e1 m\u00e1 zjednodu\u0161i\u0165 mana\u017eovanie agentov \u2013 a s GPT\u20115.3\u2011Codex m\u00e1 by\u0165 v\u00fdrazne interakt\u00edvnej\u0161ia.<\/p>\n\n\n\n

      Konkr\u00e9tne: Codex m\u00e1 poskytova\u0165 \u010dastej\u0161ie update-y o k\u013e\u00fa\u010dov\u00fdch rozhodnutiach a progrese. Namiesto \u010dakania na fin\u00e1lny output sa vie\u0161 s agentom bavi\u0165 v re\u00e1lnom \u010dase: p\u00fdta\u0165 sa, diskutova\u0165 pr\u00edstup, korigova\u0165 smerovanie. Model m\u00e1 \u201eprehov\u00e1ra\u0165\u201c cez to, \u010do rob\u00ed, reagova\u0165 na feedback a dr\u017ea\u0165 \u0165a v obraze od za\u010diatku do konca.<\/p>\n\n\n\n

      V aplik\u00e1cii sa d\u00e1 zapn\u00fa\u0165 priebe\u017en\u00e9 usmer\u0148ovanie cez nastavenie: Settings > General > Follow-up behavior<\/strong>.<\/p>\n\n\n\n

      Ako OpenAI pou\u017eilo Codex na tr\u00e9ning a nasadenie GPT\u20115.3\u2011Codex<\/h2>\n\n\n\n

      OpenAI opisuje, \u017ee posledn\u00e9 r\u00fdchle zlep\u0161enia Codexu stoja na dlhodobej\u0161\u00edch v\u00fdskumn\u00fdch projektoch naprie\u010d firmou (mesiace a\u017e roky pr\u00e1ce), ale z\u00e1rove\u0148 tieto projekty teraz Codex zr\u00fdch\u013euje. Viacer\u00ed v\u00fdskumn\u00edci a in\u017einieri vraj hovoria, \u017ee ich pr\u00e1ca je dnes z\u00e1sadne in\u00e1 ne\u017e bola pred dvomi mesiacmi.<\/p>\n\n\n\n

      Zauj\u00edmav\u00e9 je, \u017ee u\u017e ran\u00e9 verzie GPT\u20115.3\u2011Codex mali by\u0165 nato\u013eko schopn\u00e9, \u017ee t\u00edm ich pou\u017eil na zlep\u0161enie tr\u00e9ningu aj na podporu nasadenia neskor\u0161\u00edch verzi\u00ed.<\/p>\n\n\n\n

      Pr\u00edklady z v\u00fdskumu: monitoring tr\u00e9ningu, anal\u00fdza spr\u00e1vania, intern\u00e9 n\u00e1stroje<\/h3>\n\n\n\n

      V\u00fdskumn\u00fd t\u00edm pou\u017eil Codex na monitorovanie a debugovanie tr\u00e9ningov\u00e9ho behu<\/strong> pre tento release. A neostalo to pri infra probl\u00e9moch: Codex pom\u00e1hal sledova\u0165 patterny po\u010das tr\u00e9ningu, urobil hlbok\u00fa anal\u00fdzu kvality interakci\u00ed, navrhoval opravy a budoval \u201erich\u201c aplik\u00e1cie, aby \u013eudsk\u00ed v\u00fdskumn\u00edci presnej\u0161ie pochopili, ako sa spr\u00e1vanie modelu l\u00ed\u0161i oproti star\u0161\u00edm modelom.<\/p>\n\n\n\n

      Pr\u00edklady z engineeringu: optimaliz\u00e1cia harnessu a produk\u010dn\u00e9 edge cases<\/h3>\n\n\n\n

      Engineering t\u00edm pod\u013ea OpenAI pou\u017eil Codex na optimaliz\u00e1ciu a \u00fapravu harnessu<\/strong> (testovacieho\/sp\u00fa\u0161\u0165acieho r\u00e1mca) pre GPT\u20115.3\u2011Codex. Ke\u010f sa za\u010dali objavova\u0165 zvl\u00e1\u0161tne edge cases ovplyv\u0148uj\u00face pou\u017e\u00edvate\u013eov, \u013eudia v t\u00edme vyu\u017eili Codex na identifik\u00e1ciu bugov v renderovan\u00ed kontextu a odhalenie root cause pre n\u00edzke cache hit rate.<\/p>\n\n\n\n

      Po\u010das launchu m\u00e1 GPT\u20115.3\u2011Codex pom\u00e1ha\u0165 aj operat\u00edvne: dynamicky \u0161k\u00e1luje GPU clustre pod\u013ea traffic surge a pom\u00e1ha dr\u017ea\u0165 stabiln\u00fa latenciu.<\/p>\n\n\n\n

      Alpha testing: meranie produktivity \u201eper turn\u201c a anal\u00fdza session logov<\/h3>\n\n\n\n

      Po\u010das alpha testovania chcel jeden v\u00fdskumn\u00edk odmera\u0165, ko\u013eko dodato\u010dnej pr\u00e1ce model sprav\u00ed \u201ena jeden \u0165ah\u201c (per turn) a ak\u00fd je rozdiel v produktivite. GPT\u20115.3\u2011Codex navrhol nieko\u013eko jednoduch\u00fdch regex klasifik\u00e1torov<\/strong> na odhad frekvencie: (1) \u017eiadost\u00ed o upresnenie, (2) pozit\u00edvnych a negat\u00edvnych reakci\u00ed pou\u017e\u00edvate\u013ea, (3) progresu na \u00falohe. Potom ich \u0161k\u00e1lovate\u013ene spustil nad v\u0161etk\u00fdmi session logmi a vyprodukoval report so z\u00e1vermi.<\/p>\n\n\n\n

      V\u00fdsledok, ktor\u00fd OpenAI komunikuje: \u013eudia boli spokojnej\u0161\u00ed, preto\u017ee agent lep\u0161ie rozumel ich intentu a spravil viac progresu na turn, s men\u0161\u00edm po\u010dtom dopl\u0148uj\u00facich ot\u00e1zok.<\/p>\n\n\n\n

      Neintuit\u00edvne v\u00fdsledky a nov\u00e9 d\u00e1tov\u00e9 pipeline<\/h3>\n\n\n\n

      Ke\u010f\u017ee GPT\u20115.3\u2011Codex sa spr\u00e1va inak ne\u017e predchodcovia, alpha d\u00e1ta mali viac nezvy\u010dajn\u00fdch a kontraintuit\u00edvnych v\u00fdsledkov. Data scientist v t\u00edme pracoval s GPT\u20115.3\u2011Codex na vybudovan\u00ed nov\u00fdch d\u00e1tov\u00fdch pipeline a bohat\u0161ej vizualiz\u00e1cie v\u00fdsledkov, ne\u017e umo\u017e\u0148ovali \u0161tandardn\u00e9 dashboarding n\u00e1stroje. N\u00e1sledne spolu s Codexom v\u00fdsledky ko-analyzovali \u2013 Codex mal zhrn\u00fa\u0165 k\u013e\u00fa\u010dov\u00e9 insighty naprie\u010d tis\u00edckami d\u00e1tov\u00fdch bodov do troch min\u00fat.<\/p>\n\n\n\n

      OpenAI to uzatv\u00e1ra t\u00fdm, \u017ee jednotlivo s\u00fa tieto pr\u00edklady zauj\u00edmav\u00e9, ale spolu ukazuj\u00fa \u201epowerful acceleration\u201c v\u00fdskumu, engineeringu a product t\u00edmov.<\/p>\n\n\n\n

      Kyberbezpe\u010dnos\u0165: \u201eHigh capability\u201c klasifik\u00e1cia a posilnen\u00fd safety stack<\/h2>\n\n\n\n

      OpenAI tvrd\u00ed, \u017ee v posledn\u00fdch mesiacoch vid\u00ed zmyslupln\u00e9 zisky vo v\u00fdkone na kyberbezpe\u010dnostn\u00fdch \u00faloh\u00e1ch, \u010do m\u00e1 pom\u00e1ha\u0165 v\u00fdvoj\u00e1rom aj security profesion\u00e1lom. Paralelne firma pripravovala \u201estrengthened cyber safeguards\u201c na podporu defenz\u00edvneho pou\u017eitia a odolnosti ekosyst\u00e9mu.<\/p>\n\n\n\n

      GPT\u20115.3\u2011Codex je pod\u013ea OpenAI prv\u00fd model<\/strong>, ktor\u00fd klasifikuj\u00fa ako High capability<\/strong> pre kyberbezpe\u010dnostn\u00e9 \u00falohy v r\u00e1mci ich Preparedness Framework<\/strong>. Z\u00e1rove\u0148 je to prv\u00fd model, ktor\u00fd priamo tr\u00e9novali na identifik\u00e1ciu softv\u00e9rov\u00fdch zranite\u013enost\u00ed<\/strong>.<\/p>\n\n\n\n

      OpenAI dod\u00e1va d\u00f4le\u017eit\u00fa opatrnos\u0165: nemaj\u00fa definit\u00edvny d\u00f4kaz, \u017ee by model vedel automatizova\u0165 kyber\u00fatok end-to-end, ale aj tak volia prevent\u00edvny pr\u00edstup. Nasadzuj\u00fa pod\u013ea nich najkomplexnej\u0161\u00ed cybersecurity safety stack doteraz. Medzi mitig\u00e1cie uv\u00e1dzaj\u00fa:<\/p>\n\n\n\n