{"id":56756,"date":"2022-01-28T13:02:27","date_gmt":"2022-01-28T11:02:27","guid":{"rendered":"https:\/\/forklog.com\/en\/?p=56756"},"modified":"2025-09-04T08:05:04","modified_gmt":"2025-09-04T05:05:04","slug":"openai-has-created-a-less-toxic-version-of-gpt-3","status":"publish","type":"post","link":"https:\/\/u1f987.com\/en\/openai-has-created-a-less-toxic-version-of-gpt-3\/","title":{"rendered":"OpenAI has created a less toxic version of GPT-3"},"content":{"rendered":"<p>OpenAI&#8217;s AI lab has created a new version of the GPT-3 language model that produces fewer offensive expressions, misinformation and errors overall, using the <a href=\"https:\/\/ru.wikipedia.org\/wiki\/%D0%9F%D1%80%D0%BE%D0%B1%D0%BB%D0%B5%D0%BC%D0%B0_%D0%BA%D0%BE%D0%BD%D1%82%D1%80%D0%BE%D0%BB%D1%8F_%D0%B8%D1%81%D0%BA%D1%83%D1%81%D1%81%D1%82%D0%B2%D0%B5%D0%BD%D0%BD%D0%BE%D0%B3%D0%BE_%D0%B8%D0%BD%D1%82%D0%B5%D0%BB%D0%BB%D0%B5%D0%BA%D1%82%D1%83\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">the problem of AI control<\/a>.<\/p>\n<blockquote class=\"twitter-tweet\">\n<p lang=\"en\" dir=\"ltr\">We&#8217;ve trained GPT-3 to be more aligned with what humans want: The new InstructGPT models are better at following human intent than a 100x larger model, while also improving safety and truthfulness. <a href=\"https:\/\/t.co\/rKNpCDAMb2\">https:\/\/t.co\/rKNpCDAMb2<\/a><\/p>\n<p>\u2014 OpenAI (@OpenAI) <a href=\"https:\/\/twitter.com\/OpenAI\/status\/1486740126688370712?ref_src=twsrc%5Etfw\">January 27, 2022<\/a><\/p><\/blockquote>\n<p> <script async src=\"https:\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script><\/p>\n<p>To create a model named InstructGPT, researchers used reinforcement learning with human feedback. To do this, they hired 40 experts who evaluated GPT-3&#8217;s responses to a number of pre-written prompts, such as &#8220;Write a story about a wise frog named Julius&#8221; or &#8220;Write a creative advertisement for the next product to post on Facebook.&#8221;<\/p>\n<p>Responses that, in the jury&#8217;s view, more closely matched the obvious intent of the prompt author received high scores. Offensive, violent and other unacceptable results were marked as inappropriate by the experts.<\/p>\n<p>The researchers used the jury feedback as rewards in the reinforcement learning algorithm that trained InstructGPT to align responses to prompts.<\/p>\n<p>OpenAI found that users prefer InstructGPT&#8217;s responses to GPT-3 in more than 70% of cases.<\/p>\n<p>Researchers also compared versions of the new model of different sizes. They found that InstructGPT outputs with 1.3 billion parameters are preferred more than GPT-3&#8217;s outputs with 175 billion parameters. This suggests that AI control could be a straightforward way to improve language models, rather than simply increasing their size, according to the organisation.<\/p>\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u00abThis is the first time the AI-control problem has been applied to a real product\u00bb, \u2014 said one of the leaders of OpenAI&#8217;s AI-control group, Jan Leike.<\/p>\n<\/blockquote>\n<p>However, according to the researchers, InstructGPT still makes simple mistakes, sometimes producing inappropriate or nonsensical responses. For example, if given a prompt containing a lie, it will treat it as truth.<\/p>\n<p>OpenAI has made InstructGPT the default model for API users. GPT-3 remains available, but the organisation does not recommend using it.<\/p>\n<p>Earlier, OpenAI <a href=\"https:\/\/u1f987.com\/en\/news\/openai-curbs-bias-and-toxicity-in-gpt-3\">tried to soften the bias and toxicity of the base model<\/a>. Despite the progress made, the developers acknowledged a number of unresolved questions and general issues in adapting GPT-3 to society.<\/p>\n<p>In November 2021, OpenAI trained the language model <a href=\"https:\/\/u1f987.com\/en\/news\/openai-trains-language-model-to-solve-elementary-math-problems\">to solve mathematical problems<\/a>.<\/p>\n<p>In September, the lab&#8217;s researchers taught GPT-3 <a href=\"https:\/\/u1f987.com\/en\/news\/openai-develops-model-to-generate-short-extracts-from-fiction-books\">to generate brief extracts<\/a> from works of fiction.<\/p>\n<p>Subscribe to ForkLog&#8217;s AI News on Telegram: <a href=\"https:\/\/t.me\/forklogAI\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">ForkLog AI<\/a> \u2014 all the news from the world of AI!<\/p>\n","protected":false},"excerpt":{"rendered":"<p>The OpenAI AI lab has created a new version of the GPT-3 language model that produces fewer offensive expressions, misinformation and errors overall, using the problem of AI control.<\/p>\n","protected":false},"author":1,"featured_media":56757,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"select":"1","news_style_id":"1","cryptorium_level":"","_short_excerpt_text":"","creation_source":"","_metatest_mainpost_news_update":false,"footnotes":""},"categories":[3],"tags":[438,1190],"class_list":["post-56756","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-and-analysis","tag-artificial-intelligence","tag-openai"],"aioseo_notices":[],"amp_enabled":true,"views":"33","promo_type":"1","layout_type":"1","short_excerpt":"","is_update":"","_links":{"self":[{"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/posts\/56756","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/comments?post=56756"}],"version-history":[{"count":1,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/posts\/56756\/revisions"}],"predecessor-version":[{"id":56758,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/posts\/56756\/revisions\/56758"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/media\/56757"}],"wp:attachment":[{"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/media?parent=56756"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/categories?post=56756"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/u1f987.com\/en\/wp-json\/wp\/v2\/tags?post=56756"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}