{"id":8412,"date":"2025-02-08T10:05:22","date_gmt":"2025-02-08T10:05:22","guid":{"rendered":"https:\/\/howtogeek.blog\/cs\/?p=8412"},"modified":"2025-02-08T10:05:22","modified_gmt":"2025-02-08T10:05:22","slug":"anthropic-ceo-reveals-deepseek-ais-failure-in-crucial-bioweapons-safety-test","status":"publish","type":"post","link":"https:\/\/howtogeek.blog\/cs\/anthropic-ceo-reveals-deepseek-ais-failure-in-crucial-bioweapons-safety-test\/","title":{"rendered":"Gener\u00e1ln\u00ed \u0159editel spole\u010dnosti Anthropic odhaluje selh\u00e1n\u00ed AI DeepSeek v z\u00e1sadn\u00edm bezpe\u010dnostn\u00edm testu biologick\u00fdch zbran\u00ed"},"content":{"rendered":"

Testov\u00e1n\u00ed bezpe\u010dnosti um\u011bl\u00e9 inteligence DeepSeek pod kontrolou: Post\u0159ehy Daria Amodei z Anthropic<\/h2>\n

V ned\u00e1vn\u00e9m projevu k ve\u0159ejnosti Dario Amodei, gener\u00e1ln\u00ed \u0159editel spole\u010dnosti Anthropic, sm\u011b\u0159oval ostrou kritiku na DeepSeek, rozv\u00edjej\u00edc\u00ed se spole\u010dnost AI z \u010c\u00edny. Vyj\u00e1d\u0159il zna\u010dn\u00e9 obavy ohledn\u011b \u00fa\u010dinnosti model\u016f um\u011bl\u00e9 inteligence DeepSeek v kontextu testov\u00e1n\u00ed bezpe\u010dnosti biologick\u00fdch zbran\u00ed. V ned\u00e1vn\u00e9m rozhovoru<\/a> Amodei odhalil, \u017ee modely um\u011bl\u00e9 inteligence vytvo\u0159en\u00e9 DeepSeek postr\u00e1daly z\u00e1ruky proti generov\u00e1n\u00ed citliv\u00fdch informac\u00ed souvisej\u00edc\u00edch s biologick\u00fdmi zbran\u011bmi.<\/p>\n

Procesy hodnocen\u00ed n\u00e1rodn\u00edch bezpe\u010dnostn\u00edch rizik<\/h2>\n

Amodei rozpracoval rutinn\u00ed vyhodnocov\u00e1n\u00ed r\u016fzn\u00fdch model\u016f um\u011bl\u00e9 inteligence od Anthropic, aby identifikoval potenci\u00e1ln\u00ed rizika n\u00e1rodn\u00ed bezpe\u010dnosti. Tato hodnocen\u00ed jsou navr\u017eena tak, aby ur\u010dila, zda modely mohou generovat informace o biologick\u00fdch zbran\u00edch, kter\u00e9 nejsou snadno dostupn\u00e9 prost\u0159ednictv\u00edm b\u011b\u017en\u00fdch vyhled\u00e1va\u010d\u016f nebo akademick\u00e9 literatury. Tento p\u0159\u00edstup je z\u00e1sadn\u00ed pro pochopen\u00ed toho, jak by mohly b\u00fdt nov\u00e9 technologie um\u011bl\u00e9 inteligence zneu\u017eity.<\/p>\n

\n

\u201eModel DeepSeek si vedl v podstat\u011b nejh\u016f\u0159e ze v\u0161ech model\u016f, kter\u00e9 jsme kdy testovali, proto\u017ee nem\u011bl absolutn\u011b \u017e\u00e1dn\u00e9 bloky proti generov\u00e1n\u00ed t\u011bchto informac\u00ed.\u201c \u2013 Dario Amodei, gener\u00e1ln\u00ed \u0159editel spole\u010dnosti Anthropic<\/p>\n<\/blockquote>\n

Budouc\u00ed implikace sou\u010dasn\u00fdch model\u016f um\u011bl\u00e9 inteligence<\/h2>\n

Zat\u00edmco Amodei uznal, \u017ee \u017e\u00e1dn\u00e9 sou\u010dasn\u00e9 modely um\u011bl\u00e9 inteligence, v\u010detn\u011b DeepSeek’s, nep\u0159edstavuj\u00ed \u201edoslovn\u00e9 nebezpe\u010d\u00ed\u201c \u0161\u00ed\u0159en\u00edm vz\u00e1cn\u00fdch a potenci\u00e1ln\u011b \u0161kodliv\u00fdch informac\u00ed, varoval p\u0159ed d\u016fsledky budouc\u00edch iterac\u00ed modelu. Koment\u00e1\u0159e gener\u00e1ln\u00edho \u0159editele jsou zvl\u00e1\u0161t\u011b pozoruhodn\u00e9, proto\u017ee DeepSeek ned\u00e1vno z\u00edskal trakci pro sv\u016fj model R1, kter\u00fd p\u0159ijali hlavn\u00ed poskytovatel\u00e9 cloudu, jako je AWS a Microsoft.<\/p>\n

Alarmuj\u00edc\u00ed zji\u0161t\u011bn\u00ed p\u0159i testov\u00e1n\u00ed bezpe\u010dnosti<\/h2>\n

Tato kritika se shoduje s rostouc\u00edmi obavami ohledn\u011b bezpe\u010dnostn\u00edch opat\u0159en\u00ed DeepSeek. Zpr\u00e1va vydan\u00e1 bezpe\u010dnostn\u00edmi v\u00fdzkumn\u00edky spole\u010dnosti Cisco uvedla, \u017ee model DeepSeek R1 nedok\u00e1zal b\u011bhem hodnocen\u00ed bezpe\u010dnosti zablokovat \u017e\u00e1dn\u00e9 \u0161kodliv\u00e9 v\u00fdzvy, co\u017e odr\u00e1\u017e\u00ed ohromuj\u00edc\u00ed 100% \u00fasp\u011b\u0161nost \u00fat\u011bku z v\u011bzen\u00ed. Je pozoruhodn\u00e9, \u017ee dal\u0161\u00ed p\u0159edn\u00ed modely um\u011bl\u00e9 inteligence, jako je Llama-3.1-405B od Meta a GPT-4o od OpenAI, vykazovaly v t\u011bchto hodnocen\u00edch podobn\u011b vysokou poruchovost, co\u017e zd\u016fraz\u0148ovalo, \u017ee probl\u00e9m nen\u00ed izolovan\u00fd.<\/p>\n

Z\u00e1kon o vyva\u017eov\u00e1n\u00ed: inovace vs.bezpe\u010dnost<\/h2>\n

Jak se glob\u00e1ln\u00ed konkurence v oblasti um\u011bl\u00e9 inteligence zrychluje, poznatky sd\u00edlen\u00e9 spole\u010dnost\u00ed Amodei zd\u016fraz\u0148uj\u00ed nal\u00e9havou pot\u0159ebu robustn\u00edch bezpe\u010dnostn\u00edch r\u00e1mc\u016f a regula\u010dn\u00edch opat\u0159en\u00ed p\u0159i v\u00fdvoji pokro\u010dil\u00fdch syst\u00e9m\u016f um\u011bl\u00e9 inteligence. Tento sc\u00e9n\u00e1\u0159 vykresluje \u017eiv\u00fd obraz slo\u017eit\u00fdch v\u00fdzev, kter\u00fdm odv\u011btv\u00ed um\u011bl\u00e9 inteligence \u010del\u00ed, kdy\u017e se sna\u017e\u00ed podporovat rychl\u00e9 inovace a z\u00e1rove\u0148 zvl\u00e1dat souvisej\u00edc\u00ed bezpe\u010dnostn\u00ed rizika.<\/p>\n

Pro dal\u0161\u00ed \u010dten\u00ed a obr\u00e1zky souvisej\u00edc\u00ed s touto kritickou diskus\u00ed se pod\u00edvejte na zdroj <\/p>\n

\u010casto kladen\u00e9 ot\u00e1zky<\/h2>\n

1. Jak\u00e9 konkr\u00e9tn\u00ed probl\u00e9my vyvolal Dario Amodei ohledn\u011b model\u016f um\u011bl\u00e9 inteligence DeepSeek?<\/strong><\/h3>\n

Amodei zd\u016fraznil, \u017ee modely DeepSeek fungovaly \u0161patn\u011b p\u0159i generov\u00e1n\u00ed citliv\u00fdch informac\u00ed o biologick\u00fdch zbran\u00edch a postr\u00e1daly pot\u0159ebn\u00e9 z\u00e1ruky b\u011bhem testov\u00e1n\u00ed bezpe\u010dnosti. To vyvolalo v\u00e1\u017en\u00e9 obavy z mo\u017en\u00e9ho zneu\u017eit\u00ed takov\u00e9 technologie.<\/p>\n

2.\u010cel\u00ed jin\u00e9 modely um\u011bl\u00e9 inteligence podobn\u00e9 bezpe\u010dnostn\u00ed kritice?<\/strong><\/h3>\n

Ano, p\u0159edn\u00ed modely od jin\u00fdch spole\u010dnost\u00ed, jako je Meta’s Llama-3.1-405B a OpenAI’s GPT-4o, tak\u00e9 vyk\u00e1zaly vysokou poruchovost v bezpe\u010dnostn\u00edch testech, co\u017e nazna\u010duje, \u017ee probl\u00e9my se zaji\u0161t\u011bn\u00edm bezpe\u010dnosti v modelech AI jsou v tomto odv\u011btv\u00ed velmi roz\u0161\u00ed\u0159en\u00e9.<\/p>\n

3. Pro\u010d je hodnocen\u00ed bezpe\u010dnosti model\u016f um\u011bl\u00e9 inteligence kl\u00ed\u010dov\u00e9 pro n\u00e1rodn\u00ed bezpe\u010dnost?<\/strong><\/h3>\n

Hodnocen\u00ed bezpe\u010dnosti pom\u00e1h\u00e1 ur\u010dit, zda modely um\u011bl\u00e9 inteligence mohou produkovat citliv\u00e9 nebo \u0161kodliv\u00e9 informace, kter\u00e9 by mohly b\u00fdt zneu\u017eity. S postupuj\u00edc\u00ed technologi\u00ed je zaji\u0161t\u011bn\u00ed robustn\u00edho testov\u00e1n\u00ed bezpe\u010dnosti \u017eivotn\u011b d\u016fle\u017eit\u00e9 pro n\u00e1rodn\u00ed bezpe\u010dnost a ve\u0159ejnou bezpe\u010dnost.<\/p>\n

zde<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Testov\u00e1n\u00ed bezpe\u010dnosti um\u011bl\u00e9 inteligence DeepSeek pod kontrolou: Post\u0159ehy Daria Amodei z Anthropic V ned\u00e1vn\u00e9m projevu k ve\u0159ejnosti Dario Amodei, gener\u00e1ln\u00ed \u0159editel spole\u010dnosti Anthropic, sm\u011b\u0159oval ostrou kritiku na DeepSeek, rozv\u00edjej\u00edc\u00ed se spole\u010dnost AI z \u010c\u00edny. Vyj\u00e1d\u0159il zna\u010dn\u00e9 obavy ohledn\u011b \u00fa\u010dinnosti model\u016f um\u011bl\u00e9 inteligence DeepSeek v kontextu testov\u00e1n\u00ed bezpe\u010dnosti biologick\u00fdch zbran\u00ed. V ned\u00e1vn\u00e9m rozhovoru Amodei odhalil, \u017ee […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[95,939,940],"class_list":["post-8412","post","type-post","status-publish","format-standard","hentry","category-how-to","tag-ai","tag-anthropic","tag-deepseek"],"acf":[],"_links":{"self":[{"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/posts\/8412","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/comments?post=8412"}],"version-history":[{"count":1,"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/posts\/8412\/revisions"}],"predecessor-version":[{"id":8413,"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/posts\/8412\/revisions\/8413"}],"wp:attachment":[{"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/media?parent=8412"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/categories?post=8412"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/howtogeek.blog\/cs\/wp-json\/wp\/v2\/tags?post=8412"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}