{"id":483,"date":"2019-02-17T12:16:51","date_gmt":"2019-02-17T11:16:51","guid":{"rendered":"http:\/\/www.mlguru.com\/?p=483"},"modified":"2019-02-19T07:42:05","modified_gmt":"2019-02-19T06:42:05","slug":"dangerous-artificial-intelligence-by-openai","status":"publish","type":"post","link":"https:\/\/www.mlguru.com\/cs\/dangerous-artificial-intelligence-by-openai\/","title":{"rendered":"Nebezpe\u010dn\u00e1 um\u011bl\u00e1 inteligence od OpenAI?"},"content":{"rendered":"

Americk\u00e1 neziskov\u00e1 spole\u010dnost OpenAI<\/a> p\u0159ed n\u011bkolika dny na sv\u00e9m blogu zve\u0159ejnila \u010dl\u00e1nek<\/a>, ve kter\u00e9m p\u00ed\u0161e o tom, \u017ee vyvinula nov\u00fd jazykov\u00fd model, kter\u00fd je tak dobr\u00fd, \u017ee rad\u011bji nezve\u0159ejn\u00ed ani natr\u00e9novan\u00fd model, ani zdrojov\u00e9 k\u00f3dy. Pro \u0159adu m\u00e9di\u00ed (v\u010detn\u011b n\u011bkter\u00fdch seri\u00f3zn\u00edch \u010desk\u00fdch) takov\u00e9 zpr\u00e1vy funguj\u00ed jako \u010derven\u00fd hadr na b\u00fdka a okam\u017eit\u011b tak za\u010daly vznikat \u010dl\u00e1nky o tom, \u017ee byla vyvinut\u00e1 pr\u016flomov\u00e1 um\u011bl\u00e1 inteligence, kter\u00e1 je tak nebezpe\u010dn\u00e1, \u017ee se mus\u00ed dr\u017eet pod pokli\u010dkou. R\u00e1d bych uvedl v\u011bci na pravou m\u00edru.<\/strong><\/p>\n

V prvn\u00ed \u0159ad\u011b je t\u0159eba p\u0159iznat, \u017ee OpenAI se v \u010dl\u00e1nku nedopustila \u017e\u00e1dn\u00e9ho faulu a d\u016fvody, kv\u016fli kter\u00fdm necht\u011bj\u00ed zve\u0159ejnit zdrojov\u00e9 k\u00f3dy, zn\u011bj\u00ed rozumn\u011b. Je v\u0161ak evidentn\u00ed, \u017ee auto\u0159i dob\u0159e tu\u0161ili, jak budou m\u00e9dia reagovat a jde tak z jejich strany o promy\u0161len\u00fd marketingov\u00fd tah. O tom sv\u011bd\u010d\u00ed i posledn\u00ed v\u011bta \u010dl\u00e1nku, kter\u00e1 kon\u010d\u00ed slovy \u201ewe\u2019re hiring\u201c.<\/p>\n

A s \u010d\u00edm tedy v\u011bdci z OpenAI p\u0159i\u0161li? Byl vytvo\u0159en z\u0159ejm\u011b velmi dobr\u00fd jazykov\u00fd model<\/a>. Jazykov\u00fdm modelem se rozum\u00ed n\u00e1stroj, kter\u00fd na z\u00e1klad\u011b libovoln\u00e9 posloupnosti slov um\u00ed ur\u010dit pravd\u011bpodobnosti mo\u017en\u00fdch slov, kter\u00e1 by mohla n\u00e1sledovat. Pro p\u0159\u00edklad uve\u010fme posloupnost dvou slov<\/p>\n

m\u00e1ma mele<\/pre>\n
\u00dakolem jazykov\u00e9ho modelu je doplnit n\u00e1sleduj\u00edc\u00ed slovo posloupnosti. To m\u016f\u017ee b\u00fdt teoreticky jak\u00e9koliv slovo, ale jist\u011b ka\u017ed\u00fd tu\u0161\u00ed, \u017ee n\u011bkter\u00e1 slova budou pravd\u011bpodobn\u011bj\u0161\u00ed ne\u017e jin\u00e1. Pravd\u011bpodobnosti n\u00e1sleduj\u00edc\u00edch slov mohou vypadat nap\u0159\u00edklad n\u00e1sledovn\u011b:<\/p>\n\n\n\n\n\t\n\n\t\n\t\n\t\n\t\n\t

\nslovo<\/th>		pravd\u011bpodobnost<\/th>\n<\/tr>\n<\/thead>\n
maso<\/td> 0,6<\/td>\n<\/tr>\n
k\u00e1vu<\/td>0,1<\/td>\n<\/tr>\n
ko\u0159en\u00ed<\/td>0,1<\/td>\n<\/tr>\n
Emu<\/td>0,01<\/td>\n<\/tr>\n
Ostatn\u00ed slova<\/em><\/td>0,19<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n
Pokud d\u00edky jazykov\u00e9mu modelu zn\u00e1me pravd\u011bpodobnosti n\u00e1sleduj\u00edc\u00edch slov, m\u016f\u017eeme je vy\u017e\u00edt pro generov\u00e1n\u00ed \u00fapln\u011b nov\u00fdch text\u016f. Postup je n\u00e1sleduj\u00edc\u00ed:<\/p>\n
    \n
  1. Prvn\u00ed slovo zvol\u00edme sami, nebo ho vybere n\u011bjak\u00fdm n\u00e1hodn\u00fdm zp\u016fsobem.<\/li>\n
  2. Pravd\u011bpodobnosti pro n\u00e1sleduj\u00edc\u00ed slovo n\u00e1m u\u017e ur\u010duje jazykov\u00fd model. Slovo tedy vybereme n\u00e1hodn\u011b s\u00a0ohledem na tyto pravd\u011bpodobnosti. Je mo\u017en\u00e9 si to p\u0159edstavit tak, \u017ee vezmeme hrac\u00ed kostku, kter\u00e1 nem\u00e1 6 st\u011bn, ale tolik st\u011bn, kolik r\u016fzn\u00fdch slov m\u00e1me k\u00a0dispozici. Na ka\u017edou st\u011bnu takov\u00e9 kostky vlep\u00edme z\u00e1va\u017e\u00ed, jeho\u017e hmotnost bude odpov\u00eddat pravd\u011bpodobnosti dan\u00e9ho slova.<\/li>\n
  3. Onou hrac\u00ed kostkou si hod\u00edme a zap\u00ed\u0161eme si slovo, kter\u00e9 n\u00e1m padlo. T\u00edm n\u00e1m vznikne posloupnost o jedno slovo del\u0161\u00ed. Na jej\u00edm z\u00e1klad\u011b podle jazykov\u00e9ho modelu uprav\u00edme pravd\u011bpodobnosti pro n\u00e1sleduj\u00edc\u00ed slovo a cel\u00fd postup opakujeme od bodu 2.<\/li>\n<\/ol>\n
    Kl\u00ed\u010dem k \u00fasp\u011bchu a k vytvo\u0159en\u00ed smyslupln\u00e9 posloupnosti slov je tedy kvalitn\u00ed jazykov\u00fd model. Na to existuj\u00ed u\u017e lety prov\u011b\u0159en\u00e9 postupy. Nejprve pot\u0159ebujeme dostate\u010dn\u011b velkou kolekci textov\u00fdch soubor\u016f, kter\u00e1 slou\u017e\u00ed jako vzorov\u00e9 texty. V oboru strojov\u00e9ho u\u010den\u00ed se takov\u00e9mu datov\u00e9mu souboru \u0159\u00edk\u00e1 tr\u00e9novac\u00ed data. Pokud se v t\u011bchto tr\u00e9novac\u00edch datech bude po slovech m\u00e1ma mele<\/code> vyskytovat maso<\/code> \u010dast\u011bji, ne\u017e jin\u00e1 slova, budeme cht\u00edt, aby jazykov\u00fd model posloupnosti m\u00e1ma mele maso<\/code> p\u0159i\u0159adil velkou pravd\u011bpodobnost.<\/p>\n
    Na druhou stranu ale nechceme, ale si jazykov\u00fd model jen zapamatoval \u010dast\u00e9 posloupnosti. Chceme, aby byl schopn\u00fd zobec\u0148ovat i na posloupnosti, kter\u00e9 se v tr\u00e9novac\u00edch datech nevyskytly. K tomu slou\u017e\u00ed algoritmy strojov\u00e9ho u\u010den\u00ed.<\/p>\n
    Podle \u010dl\u00e1nku OpenAI byly pro strojov\u00e9 u\u010den\u00ed pou\u017eity \u201estar\u00e9 zn\u00e1m\u00e9\u201c algoritmy (velmi podobn\u00e9 nap\u0159\u00edklad tomu, kter\u00fd byl slou\u017eit pro generov\u00e1n\u00ed b\u00e1sn\u00ed<\/a>), jen tr\u00e9novac\u00ed data a kapacita pam\u011bti modelu byla o mnoho \u0159\u00e1d\u016f v\u011bt\u0161\u00ed ne\u017e d\u0159\u00edve. To bylo umo\u017en\u011bno d\u00edky velmi v\u00fdkonn\u00e9 v\u00fdpo\u010detn\u00ed infrastruktu\u0159e, kter\u00e1 b\u011b\u017en\u00fdm smrteln\u00edk\u016fm mimo \u0161pi\u010dkov\u00e1 AI centra nen\u00ed dostupn\u00e1.<\/p>\n
    V \u010dem tedy spo\u010d\u00edv\u00e1 nebezpe\u010d\u00ed takov\u00e9ho modelu? V \u017e\u00e1dn\u00e9m p\u0159\u00edpad\u011b se nejedn\u00e1 o n\u011bjakou formu samostatn\u011b uva\u017euj\u00edc\u00ed um\u011bl\u00e9 inteligence. St\u00e1le jsou to jen n\u00e1hodn\u00e9 procesy a statistiky po\u010d\u00edtan\u00e9 z tr\u00e9novac\u00edch dat. D\u00edky jejich slo\u017eitosti je v\u0161ak mo\u017en\u00e9 generovat texty, kter\u00e9 jsou velice podobn\u00e9 text\u016fm psan\u00fdm lidmi. Zneu\u017e\u00edt je lze nap\u0159\u00edklad pro strojov\u00e9 generov\u00e1n\u00ed fake news nebo koment\u00e1\u0159\u016f v diskuz\u00edch. S pomoc\u00ed t\u011bchto p\u0159\u00edstup\u016f je mo\u017en\u00e9 vygenerovat libovoln\u00e9 mno\u017estv\u00ed r\u016fznorod\u00e9ho textu, kter\u00e9 nebude st\u00e1t v\u00edc ne\u017e trochu elektrick\u00e9 energie.<\/p>\n
    P\u0159itom texty nemus\u00ed b\u00fdt jen bezduch\u00fdm bl\u00e1bolen\u00edm. Zam\u011b\u0159en\u00ed text\u016f je v\u017edy d\u00e1no tr\u00e9novac\u00edmi daty. Pokud se pro tr\u00e9nov\u00e1n\u00ed pou\u017eij\u00ed ideov\u011b zabarven\u00e9 propagandistick\u00e9 texty, v\u00fdsledek bude k nerozezn\u00e1n\u00ed od lidmi psan\u00e9ho obsahu na n\u011bkter\u00fdch fake news serverech.<\/p>\n
    Sami auto\u0159i \u010dl\u00e1nku p\u00ed\u0161\u00ed, \u017ee replikace jejich v\u00fdsledk\u016f nebude velk\u00fd probl\u00e9m. P\u0159esto\u017ee byly zve\u0159ejn\u011bny pouze zdrojov\u00e9 k\u00f3dy men\u0161\u00edho modelu, pou\u017eit\u00e9 principy jsou ve v\u011bdeck\u00e9 komunit\u011b dob\u0159e zn\u00e1my a jeho roz\u0161\u00ed\u0159en\u00ed je pro znal\u00e9ho \u010dlov\u011bka snadn\u00e9. Jedin\u00fdm z\u00e1sadn\u00edm omezen\u00edm je obrovsk\u00e1 v\u00fdpo\u010detn\u00ed n\u00e1ro\u010dnost. Ta se v\u0161ak t\u00fdk\u00e1 jen procesu vytv\u00e1\u0159en\u00ed modelu. Jakmile model n\u011bkdo vytvo\u0159\u00ed a zve\u0159ejn\u00ed, jeho pou\u017eit\u00ed u\u017e je velmi snadn\u00e9 i s vyu\u017eit\u00edm \u0159\u00e1dov\u011b men\u0161\u00edho v\u00fdpo\u010detn\u00edho v\u00fdkonu. Nav\u00edc z Moorova z\u00e1kona<\/a> v\u00edme, \u017ee v\u00fdpo\u010detn\u00ed v\u00fdkon po\u010d\u00edta\u010d\u016f roste v \u010dase p\u0159ibli\u017en\u011b exponenci\u00e1ln\u011b a nebude tedy trvat dlouho, ne\u017e si takov\u00fd gener\u00e1tor bude moci ka\u017ed\u00fd z n\u00e1s natr\u00e9novat na sv\u00e9m mobilu.<\/p>\n
    Ze strany OpenAI tedy o \u017e\u00e1dn\u00fd velk\u00fd pr\u016flom nejde. Nav\u00edc by bylo naivn\u00ed si myslet, \u017ee u\u017e se takov\u00fdmito generovan\u00fdmi texty n\u00e1\u0161 medi\u00e1ln\u00ed prosto d\u00e1vno nepln\u00ed. Jen se t\u00edm auto\u0159i t\u011bchto gener\u00e1tor\u016f text\u016f ve\u0159ejn\u011b nechlub\u00ed.<\/p>","protected":false},"excerpt":{"rendered":"
    Americk\u00e1 neziskov\u00e1 spole\u010dnost OpenAI p\u0159ed n\u011bkolika dny na sv\u00e9m blogu zve\u0159ejnila \u010dl\u00e1nek, ve kter\u00e9m p\u00ed\u0161e o tom, \u017ee vyvinula nov\u00fd jazykov\u00fd model, kter\u00fd je tak dobr\u00fd, \u017ee rad\u011bji nezve\u0159ejn\u00ed ani natr\u00e9novan\u00fd model, ani zdrojov\u00e9 k\u00f3dy. Pro \u0159adu m\u00e9di\u00ed (v\u010detn\u011b n\u011bkter\u00fdch seri\u00f3zn\u00edch \u010desk\u00fdch) takov\u00e9 zpr\u00e1vy funguj\u00ed jako \u010derven\u00fd hadr na b\u00fdka a okam\u017eit\u011b tak za\u010daly vznikat […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_exactmetrics_skip_tracking":false,"_exactmetrics_sitenote_active":false,"_exactmetrics_sitenote_note":"","_exactmetrics_sitenote_category":0,"footnotes":""},"categories":[10,14,12],"tags":[],"class_list":["post-483","post","type-post","status-publish","format-standard","hentry","category-nlp","category-strojove-uceni","category-umela-inteligence"],"_links":{"self":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts\/483","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/comments?post=483"}],"version-history":[{"count":13,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts\/483\/revisions"}],"predecessor-version":[{"id":501,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/posts\/483\/revisions\/501"}],"wp:attachment":[{"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/media?parent=483"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/categories?post=483"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mlguru.com\/cs\/wp-json\/wp\/v2\/tags?post=483"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}