{"id":1093,"date":"2026-07-17T04:30:32","date_gmt":"2026-07-16T19:30:32","guid":{"rendered":"https:\/\/itexplore.org\/jp\/columns\/ai-agents-quantum-computing-llm-safety-tech-trends\/"},"modified":"2026-07-17T04:30:32","modified_gmt":"2026-07-16T19:30:32","slug":"ai-agents-quantum-computing-llm-safety-tech-trends","status":"publish","type":"post","link":"https:\/\/itexplore.org\/jp\/columns\/ai-agents-quantum-computing-llm-safety-tech-trends\/","title":{"rendered":"AI\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u3001\u91cf\u5b50\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u3001LLM\u5b89\u5168\u6027\uff1a\u6700\u65b0\u6280\u8853\u52d5\u5411"},"content":{"rendered":"<p>\u672c\u65e5\u306e\u6ce8\u76eeAI\u30fb\u30c6\u30c3\u30af\u30cb\u30e5\u30fc\u30b9\u3092\u3001\u5c02\u9580\u7684\u306a\u5206\u6790\u3068\u5171\u306b\u304a\u5c4a\u3051\u3057\u307e\u3059\u3002<\/p>\n<div class=\"wp-block-vk-blocks-alert vk_alert alert alert-warning has-alert-icon\">\n<div class=\"vk_alert_icon\">\n<div class=\"vk_alert_icon_icon\"><i class=\"fa-solid fa-triangle-exclamation\" aria-hidden=\"true\"><\/i><\/div>\n<div class=\"vk_alert_icon_text\"><span>Warning<\/span><\/div>\n<\/div>\n<div class=\"vk_alert_content\">\n<p>\u3053\u306e\u8a18\u4e8b\u306fAI\u306b\u3088\u3063\u3066\u81ea\u52d5\u751f\u6210\u30fb\u5206\u6790\u3055\u308c\u305f\u3082\u306e\u3067\u3059\u3002AI\u306e\u6027\u8cea\u4e0a\u3001\u4e8b\u5b9f\u8aa4\u8a8d\u304c\u542b\u307e\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308b\u305f\u3081\u3001\u91cd\u8981\u306a\u5224\u65ad\u3092\u4e0b\u3059\u969b\u306f\u5fc5\u305a\u30ea\u30f3\u30af\u5148\u306e\u4e00\u6b21\u30bd\u30fc\u30b9\u3092\u3054\u78ba\u8a8d\u304f\u3060\u3055\u3044\u3002<\/p>\n<\/div>\n<\/div>\n<div class=\"wp-block-group\" style=\"margin-top:40px;margin-bottom:40px\">\n<h2 class=\"wp-block-heading\">\u91cf\u5b50\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30fc\u304c\u6838\u878d\u5408\u306e\u4e3b\u8981\u8ab2\u984c\u300c\u71c3\u6599\u751f\u7523\u300d\u306b\u53d6\u308a\u7d44\u3080<\/h2>\n<ul>\n<li><strong>\u539f\u984c:<\/strong> A quantum computer has just tackled one of the big problems of nuclear fusion: producing its own fuel<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">\u5c02\u9580\u30a2\u30ca\u30ea\u30b9\u30c8\u306e\u5206\u6790<\/h3>\n<div class=\"ai-summary-content\">\n<p><strong>\u30aa\u30fc\u30af\u30ea\u30c3\u30b8\u56fd\u7acb\u7814\u7a76\u6240<\/strong>\u3001<strong>\u30af\u30ea\u30fc\u30d6\u30e9\u30f3\u30c9\u30af\u30ea\u30cb\u30c3\u30af<\/strong>\u3001<strong>IBM<\/strong>\u3001\u304a\u3088\u3073<strong>\u30df\u30b7\u30ac\u30f3\u5dde\u7acb\u5927\u5b66<\/strong>\u306e\u7814\u7a76\u8005\u305f\u3061\u306f\u3001\u6838\u878d\u5408\u7089\u306e\u5c06\u6765\u306e\u5019\u88dc\u3067\u3042\u308b\u6eb6\u878d\u5869<strong>FLiBe<\/strong>\u3068\u30c8\u30ea\u30c1\u30a6\u30e0\u306e\u76f8\u4e92\u4f5c\u7528\u3092\u7814\u7a76\u3059\u308b\u305f\u3081\u306b\u3001\u521d\u3081\u3066\u91cf\u5b50\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u3092\u4f7f\u7528\u3057\u307e\u3057\u305f\u3002\u3053\u306e\u753b\u671f\u7684\u306a\u7814\u7a76\u306f\u3001\u6838\u878d\u5408\u306e\u4e3b\u8981\u306a\u8ab2\u984c\u306e\u4e00\u3064\u3067\u3042\u308b\u3001\u53cd\u5fdc\u5668\u5185\u3067\u81ea\u8eab\u306e\u71c3\u6599\uff08\u30c8\u30ea\u30c1\u30a6\u30e0\uff09\u3092\u751f\u7523\u3059\u308b\u3068\u3044\u3046\u554f\u984c\u306b\u53d6\u308a\u7d44\u3080\u3082\u306e\u3067\u3059\u3002<\/p>\n<p>\u6838\u878d\u5408\u306f\u5927\u91cf\u306e\u30af\u30ea\u30fc\u30f3\u30a8\u30cd\u30eb\u30ae\u30fc\u3092\u7d04\u675f\u3057\u307e\u3059\u304c\u3001\u305d\u306e\u4e3b\u8981\u71c3\u6599\u3067\u3042\u308b\u30c8\u30ea\u30c1\u30a6\u30e0\u306f\u5730\u7403\u4e0a\u306b\u81ea\u7136\u306b\u307b\u3068\u3093\u3069\u5b58\u5728\u3057\u307e\u305b\u3093\u3002\u3053\u306e\u7814\u7a76\u3067\u306f\u3001<strong>FLiBe<\/strong>\uff08\u30d5\u30c3\u5316\u30ea\u30c1\u30a6\u30e0\u3068\u30d5\u30c3\u5316\u30d9\u30ea\u30ea\u30a6\u30e0\u306e\u6df7\u5408\u7269\uff09\u304c\u3001\u6838\u878d\u5408\u4e2d\u306b\u653e\u51fa\u3055\u308c\u308b\u9ad8\u30a8\u30cd\u30eb\u30ae\u30fc\u4e2d\u6027\u5b50\u304c\u30ea\u30c1\u30a6\u30e0\u540c\u4f4d\u4f53\u306b\u885d\u7a81\u3059\u308b\u3053\u3068\u3067\u30c8\u30ea\u30c1\u30a6\u30e0\u3092\u751f\u6210\u3059\u308b\u300c\u30d6\u30e9\u30f3\u30b1\u30c3\u30c8\u300d\u5c64\u306e\u4e00\u90e8\u3068\u3057\u3066\u6a5f\u80fd\u3059\u308b\u53ef\u80fd\u6027\u3092\u63a2\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u7814\u7a76\u30c1\u30fc\u30e0\u306f\u3001\u30c8\u30ea\u30c1\u30a6\u30e0\u306e\u6709\u7121\u306b\u304b\u304b\u308f\u3089\u305a\u3001<strong>FLiBe<\/strong>\u306e9\u3064\u306e\u5206\u5b50\u69cb\u6210\u3092\u5206\u6790\u3057\u3001\u30a8\u30cd\u30eb\u30ae\u30fc\u3001\u96fb\u5b50\u69cb\u9020\u3001\u7d50\u5408\u5b89\u5b9a\u6027\u306a\u3069\u306e\u7279\u6027\u3092\u8a08\u7b97\u3057\u307e\u3057\u305f\u3002\u3053\u306e\u4f5c\u696d\u306f\u3001<strong>IBM<\/strong>\u306e\u91cf\u5b50\u30cf\u30fc\u30c9\u30a6\u30a7\u30a2\u3068<strong>\u30aa\u30fc\u30af\u30ea\u30c3\u30b8\u56fd\u7acb\u7814\u7a76\u6240<\/strong>\u306e\u9ad8\u6027\u80fd\u53e4\u5178\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u30ea\u30bd\u30fc\u30b9\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u91cf\u5b50\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u5229\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u3053\u306e\u5b9f\u9a13\u306e\u91cd\u8981\u6027\u306f\u3001\u5358\u4e00\u306e\u5b8c\u74a7\u306a\u7d44\u6210\u3092\u767a\u898b\u3057\u305f\u3053\u3068\u3067\u306f\u306a\u304f\u3001\u91cf\u5b50\u30d7\u30ed\u30bb\u30c3\u30b5\u304c\u6838\u878d\u5408\u306b\u95a2\u9023\u3059\u308b\u6750\u6599\u306e\u7814\u7a76\u306b\u7d44\u307f\u8fbc\u307e\u308c\u308b\u53ef\u80fd\u6027\u3092\u793a\u3057\u305f\u3053\u3068\u306b\u3042\u308a\u307e\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u96fb\u5b50\u9593\u306e\u76f8\u4e92\u4f5c\u7528\u304c\u5f93\u6765\u306e\u53e4\u5178\u7684\u306a\u65b9\u6cd5\u3067\u306f\u8907\u96d1\u3059\u304e\u308b\u5206\u91ce\u306b\u304a\u3044\u3066\u3001\u8a2d\u8a08\u30b5\u30a4\u30af\u30eb\u3092\u52a0\u901f\u3067\u304d\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<p>\u91cf\u5b50\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u306f\u307e\u3060\u6838\u878d\u5408\u306e\u71c3\u6599\u4f9b\u7d66\u554f\u984c\u3092\u5b8c\u5168\u306b\u89e3\u6c7a\u3057\u3066\u3044\u307e\u305b\u3093\u304c\u3001\u3053\u306e\u65b0\u3057\u3044\u5b9f\u9a13\u306f\u3001\u5c06\u6765\u7684\u306b\u53cd\u5fdc\u5668\u5185\u3067\u71c3\u6599\u3092\u751f\u7523\u3059\u308b\u6750\u6599\u3092\u539f\u5b50\u30ec\u30d9\u30eb\u3067\u7814\u7a76\u3059\u308b\u305f\u3081\u306e\u9053\u3092\u958b\u304d\u307e\u3059\u3002\u3053\u308c\u306f\u3001\u6838\u878d\u5408\u30a8\u30cd\u30eb\u30ae\u30fc\u306e\u5b9f\u73fe\u306b\u5411\u3051\u305f\u91cd\u8981\u306a\u4e00\u6b69\u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n<p>\ud83d\udc49 <strong><a href=\"https:\/\/es.gizmodo.com\/un-ordenador-cuantico-acaba-de-abordar-uno-de-los-grandes-problemas-de-la-fusion-nuclear-producir-su-propio-combustible-2000246230\" target=\"_blank\" rel=\"noopener\">Gizmodo en Espa\u00f1ol \u3067\u8a18\u4e8b\u5168\u6587\u3092\u8aad\u3080<\/a><\/strong><\/p>\n<\/div>\n<ul>\n<li><strong>\u8981\u70b9:<\/strong> Quantum computing offers a promising new tool for accelerating the design and understanding of materials critical for producing tritium fuel in future nuclear fusion reactors, though it's a complementary tool to classical supercomputing.<\/li>\n<li><strong>\u8457\u8005:<\/strong> Thomas Handley<\/li>\n<\/ul>\n<blockquote class=\"wp-block-quote\"><p><span>English Summary:<\/span><\/p>\n<p>Researchers from <strong>Oak Ridge National Laboratory<\/strong>, <strong>Cleveland Clinic<\/strong>, <strong>IBM<\/strong>, and <strong>Michigan State University<\/strong> have for the first time used quantum computing to study how tritium interacts with <strong>FLiBe<\/strong>, a molten salt candidate for future fusion reactors. This groundbreaking work addresses one of the major challenges of nuclear fusion: producing its own fuel (tritium) within the reactor.<\/p>\n<p>Nuclear fusion promises vast amounts of clean energy, but tritium, one of its main fuels, is scarcely found naturally on Earth. The study explores <strong>FLiBe<\/strong>, a mixture of lithium fluoride and beryllium fluoride, as part of a 'breeding blanket' layer that could produce tritium when high-energy neutrons from fusion impact lithium isotopes within it.<\/p>\n<p>The team analyzed nine molecular configurations of <strong>FLiBe<\/strong>, both with and without tritium, to calculate properties such as energy, electronic structure, and bond stability. This was achieved using a hybrid quantum computing approach, combining <strong>IBM<\/strong>'s quantum hardware with high-performance classical computing resources from <strong>Oak Ridge National Laboratory<\/strong>.<\/p>\n<p>The significance of this experiment lies not in finding a single perfect composition, but in demonstrating that quantum processors can now be integrated into the study of materials relevant to fusion. This could accelerate design cycles in a field where electron interactions can become too complex for certain conventional classical methods.<\/p>\n<p>While quantum computing has not yet fully solved the fusion fuel supply problem, this new experiment opens a path to study, atom by atom, the materials that will one day need to produce fuel within the reactor itself. This marks a crucial step towards the realization of fusion energy.<\/p>\n<\/blockquote>\n<\/div>\n<div class=\"wp-block-group\" style=\"margin-top:40px;margin-bottom:40px\">\n<h2 class=\"wp-block-heading\">\u30b8\u30e3\u30fc\u30ca\u30ea\u30b9\u30c8\u304cClaude\u306b\u5225\u306eAI\u306e\u8a13\u7df4\u30fb\u8a55\u4fa1\u30fb\u4fee\u6b63\u3092\u4efb\u305b\u308b\u5b9f\u9a13\uff1a\u5c02\u9580\u30e2\u30c7\u30eb\u4f5c\u6210\u306e\u6c11\u4e3b\u5316<\/h2>\n<ul>\n<li><strong>\u539f\u984c:<\/strong> A journalist let Claude train, evaluate, and correct another artificial intelligence for a week. The experiment demonstrates that creating specialized models is no longer an exclusive task of OpenAI, Google, or Anthropic<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">\u5c02\u9580\u30a2\u30ca\u30ea\u30b9\u30c8\u306e\u5206\u6790<\/h3>\n<div class=\"ai-summary-content\">\n<p>\u30b8\u30e3\u30fc\u30ca\u30ea\u30b9\u30c8\u306e<strong>Will Knight<\/strong>\u6c0f\u304c\u3001<strong>Claude<\/strong>\u3001<strong>AutoResearch<\/strong>\u3001\u304a\u3088\u3073<strong>Prime Intellect<\/strong>\u3068\u3044\u3063\u305f\u65e2\u5b58\u306eAI\u30e2\u30c7\u30eb\u3068\u30c4\u30fc\u30eb\u3092\u6d3b\u7528\u3057\u3001\u5225\u306eAI\u3092\u8a13\u7df4\u30fb\u8a55\u4fa1\u30fb\u4fee\u6b63\u3059\u308b\u5b9f\u9a13\u3092\u884c\u3044\u307e\u3057\u305f\u3002\u3053\u306e\u5b9f\u9a13\u306f\u3001\u7279\u5b9a\u306e\u30bf\u30b9\u30af\u306b\u7279\u5316\u3057\u305fAI\u30e2\u30c7\u30eb\u306e\u69cb\u7bc9\u304c\u3001\u3082\u306f\u3084<strong>OpenAI<\/strong>\u3001<strong>Google<\/strong>\u3001<strong>Anthropic<\/strong>\u306e\u3088\u3046\u306a\u5927\u624b\u4f01\u696d\u3060\u3051\u306e\u5c02\u58f2\u7279\u8a31\u3067\u306f\u306a\u3044\u3053\u3068\u3092\u793a\u5506\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p><strong>Knight<\/strong>\u6c0f\u306f\u3001\u79d1\u5b66\u7814\u7a76\u3092\u9078\u629e\u3057\u8981\u7d04\u3059\u308b\u5c02\u9580\u30e2\u30c7\u30eb\u3092\u69cb\u7bc9\u3059\u308b\u3053\u3068\u3092\u76ee\u6307\u3057\u307e\u3057\u305f\u3002\u5f7c\u306f\u3001<strong>Andrej Karpathy<\/strong>\u6c0f\uff08<strong>OpenAI<\/strong>\u5171\u540c\u5275\u8a2d\u8005\u3001\u5143<strong>Tesla<\/strong> AI\u8cac\u4efb\u8005\u3001\u73fe<strong>Anthropic<\/strong>\u7814\u7a76\u8005\uff09\u304c\u4f5c\u6210\u3057\u305f\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u30d7\u30ed\u30b8\u30a7\u30af\u30c8<strong>AutoResearch<\/strong>\u3092\u4f7f\u7528\u3057\u3001<strong>Claude<\/strong>\u306b\u5c0f\u3055\u306a\u8a00\u8a9e\u30e2\u30c7\u30eb\u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u30b3\u30fc\u30c9\u3092\u4fee\u6b63\u3055\u305b\u307e\u3057\u305f\u3002<\/p>\n<p><strong>AutoResearch<\/strong>\u306f\u3001AI\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u304c\u30e2\u30c7\u30eb\u306e\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3001\u30aa\u30d7\u30c6\u30a3\u30de\u30a4\u30b6\u30fc\u3001\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba\u306a\u3069\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u5909\u66f4\u3057\u3001\u5404\u5909\u66f4\u5f8c\u306b5\u5206\u9593\u306e\u30c6\u30b9\u30c8\u3092\u5b9f\u884c\u3057\u3066\u7d50\u679c\u3092\u6e2c\u5b9a\u3057\u3001\u6539\u5584\u304c\u898b\u3089\u308c\u308c\u3070\u5909\u66f4\u3092\u4fdd\u6301\u3059\u308b\u3068\u3044\u3046\u4ed5\u7d44\u307f\u3067\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u4eba\u9593\u304c\u5404\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u624b\u52d5\u3067\u8abf\u6574\u3057\u305f\u308a\u3001\u3059\u3079\u3066\u306e\u30c6\u30b9\u30c8\u3092\u5b9f\u884c\u3057\u305f\u308a\u3059\u308b\u5fc5\u8981\u304c\u306a\u304f\u306a\u308a\u307e\u3059\u3002<\/p>\n<p>\u3055\u3089\u306b\u3001<strong>Knight<\/strong>\u6c0f\u306f<strong>Prime Intellect<\/strong>\u306e\u30d7\u30e9\u30c3\u30c8\u30d5\u30a9\u30fc\u30e0\u3092\u5229\u7528\u3057\u3066\u3001<strong>Frontier_Paper_Curator<\/strong>\u3068\u3044\u3046\u30e2\u30c7\u30eb\u3092\u958b\u767a\u3057\u307e\u3057\u305f\u3002\u3053\u306e\u30e2\u30c7\u30eb\u306f\u3001<strong>Claude<\/strong>\u304c\u8ffd\u52a0\u306e\u7814\u7a76\u3092\u691c\u7d22\u3057\u3001\u5408\u6210\u4f8b\u3092\u751f\u6210\u3057\u3001\u5f37\u5316\u5b66\u7fd2\u3092\u901a\u3058\u3066\u5fdc\u7b54\u3092\u8abf\u6574\u3059\u308b\u3053\u3068\u3067\u3001\u95a2\u9023\u3059\u308b\u7814\u7a76\u3092\u9078\u629e\u3057\u8981\u7d04\u3059\u308b\u80fd\u529b\u3092\u5b66\u7fd2\u3057\u307e\u3057\u305f\u3002<\/p>\n<p>\u3053\u306e\u5b9f\u9a13\u306f\u3001AI\u304c\u81ea\u3089\u3092\u518d\u5e30\u7684\u306b\u6539\u5584\u3059\u308b\u300c\u81ea\u5df1\u6539\u5584\u300d\u306eSF\u7684\u306a\u30b7\u30ca\u30ea\u30aa\u3068\u306f\u7570\u306a\u308a\u307e\u3059\u304c\u3001\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u306e\u53cd\u5fa9\u7684\u306a\u90e8\u5206\u3092\u81ea\u52d5\u5316\u3059\u308b\u3053\u3068\u3067\u3001\u5c0f\u898f\u6a21\u306a\u30c1\u30fc\u30e0\u3067\u3082\u624b\u52d5\u3067\u306f\u4e0d\u53ef\u80fd\u3060\u3063\u305f\u591a\u304f\u306e\u4ee3\u66ff\u6848\u3092\u63a2\u6c42\u3067\u304d\u308b\u53ef\u80fd\u6027\u3092\u793a\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306f\u3001\u6cd5\u5f8b\u6587\u66f8\u306e\u5206\u6790\u3001\u5728\u5eab\u7ba1\u7406\u3001\u7523\u696d\u4e0a\u306e\u6b20\u9665\u691c\u51fa\u306a\u3069\u3001\u7279\u5b9a\u306e\u30bf\u30b9\u30af\u306b\u7279\u5316\u3057\u305f\u6570\u5343\u3082\u306e\u5c0f\u898f\u6a21\u30e2\u30c7\u30eb\u306e\u672a\u6765\u3092\u793a\u5506\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\ud83d\udc49 <strong><a href=\"https:\/\/es.gizmodo.com\/un-periodista-dejo-que-claude-entrenara-evaluara-y-corrigiera-otra-inteligencia-artificial-durante-una-semana-el-experimento-demuestra-que-crear-modelos-especializados-ya-no-es-una-tarea-exclusiva-d-2000246258\" target=\"_blank\" rel=\"noopener\">Gizmodo en Espa\u00f1ol \u3067\u8a18\u4e8b\u5168\u6587\u3092\u8aad\u3080<\/a><\/strong><\/p>\n<\/div>\n<ul>\n<li><strong>\u8981\u70b9:<\/strong> AI agents like Claude, combined with open-source tools, can automate significant portions of AI model training and evaluation, democratizing the creation of specialized AI models beyond major tech companies.<\/li>\n<li><strong>\u8457\u8005:<\/strong> Mart\u00edn Nicol\u00e1s Parolari<\/li>\n<\/ul>\n<blockquote class=\"wp-block-quote\"><p><span>English Summary:<\/span><\/p>\n<p>Journalist <strong>Will Knight<\/strong> conducted an experiment where he used existing AI models and tools like <strong>Claude<\/strong>, <strong>AutoResearch<\/strong>, and <strong>Prime Intellect<\/strong> to train, evaluate, and correct another AI. This experiment suggests that building specialized AI models is no longer an exclusive task for major companies like <strong>OpenAI<\/strong>, <strong>Google<\/strong>, or <strong>Anthropic<\/strong>.<\/p>\n<p><strong>Knight<\/strong> aimed to build a specialized model capable of selecting and summarizing scientific research. He utilized <strong>AutoResearch<\/strong>, an open-source project created by <strong>Andrej Karpathy<\/strong> (co-founder of <strong>OpenAI<\/strong>, former head of AI at <strong>Tesla<\/strong>, and current researcher at <strong>Anthropic<\/strong>), allowing <strong>Claude<\/strong> to modify the training code of a small language model.<\/p>\n<p><strong>AutoResearch<\/strong> enables an AI agent to modify parameters such as model architecture, optimizer, and batch size, running a five-minute test after each change to measure results and retain modifications if improvements are observed. This eliminates the need for humans to manually adjust every parameter or run all tests.<\/p>\n<p>Furthermore, <strong>Knight<\/strong> used <strong>Prime Intellect<\/strong>'s platform to develop a model named <strong>Frontier_Paper_Curator<\/strong>. This model learned to locate and summarize relevant research, with <strong>Claude<\/strong> assisting in finding more studies, producing synthetic examples, and adjusting responses through reinforcement learning.<\/p>\n<p>While this experiment does not represent the sci-fi scenario of recursive self-improvement, it demonstrates that automating the repetitive parts of training can allow small teams to explore far more alternatives than they could manually. This points to a future with thousands of small, specialized models created for very specific tasks, such as analyzing legal documents, managing inventories, or detecting industrial faults.<\/p>\n<\/blockquote>\n<\/div>\n<div class=\"wp-block-group\" style=\"margin-top:40px;margin-bottom:40px\">\n<h2 class=\"wp-block-heading\">GPT-Red\u306e\u7d39\u4ecb\uff1aOpenAI\u304c\u30e2\u30c7\u30eb\u3092\u3088\u308a\u5b89\u5168\u306b\u3059\u308b\u305f\u3081\u306b\u69cb\u7bc9\u3057\u305fLLM\u30b9\u30fc\u30d1\u30fc\u30cf\u30c3\u30ab\u30fc<\/h2>\n<ul>\n<li><strong>\u539f\u984c:<\/strong> Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">\u5c02\u9580\u30a2\u30ca\u30ea\u30b9\u30c8\u306e\u5206\u6790<\/h3>\n<div class=\"ai-summary-content\">\n<p><strong>OpenAI<\/strong>\u306f\u3001\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\uff08<strong>LLM<\/strong>\uff09\u306e\u5b89\u5168\u6027\u3092\u5411\u4e0a\u3055\u305b\u308b\u305f\u3081\u306b\u3001<strong>GPT-Red<\/strong>\u3068\u3044\u3046\u540d\u306e\u300c\u30b9\u30fc\u30d1\u30fc\u30cf\u30c3\u30ab\u30fc\u300d<strong>LLM<\/strong>\u3092\u958b\u767a\u3057\u307e\u3057\u305f\u3002\u3053\u306e\u30b7\u30b9\u30c6\u30e0\u306f\u3001\u60aa\u610f\u306e\u3042\u308b\u30e6\u30fc\u30b6\u30fc\u304c<strong>LLM<\/strong>\u306e\u30ac\u30fc\u30c9\u30ec\u30fc\u30eb\u3092\u56de\u907f\u3057\u3001\u6709\u5bb3\u306a\u30b3\u30f3\u30c6\u30f3\u30c4\u3092\u751f\u6210\u3055\u305b\u308b\u53ef\u80fd\u6027\u306e\u3042\u308b\u8106\u5f31\u6027\u3084\u300c\u30b8\u30a7\u30a4\u30eb\u30d6\u30ec\u30a4\u30af\u300d\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u7279\u5b9a\u3059\u308b\u305f\u3081\u306b\u8a2d\u8a08\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<p><strong>GPT-Red<\/strong>\u306f\u3001\u65e2\u5b58\u306e<strong>LLM<\/strong>\u306e\u5f31\u70b9\u3092\u7a4d\u6975\u7684\u306b\u63a2\u3057\u51fa\u3059\u3053\u3068\u3067\u3001<strong>OpenAI<\/strong>\u304c\u30e2\u30c7\u30eb\u306e\u5805\u7262\u6027\u3092\u9ad8\u3081\u3001\u502b\u7406\u7684\u30ac\u30a4\u30c9\u30e9\u30a4\u30f3\u3092\u9075\u5b88\u3067\u304d\u308b\u3088\u3046\u306b\u652f\u63f4\u3057\u307e\u3059\u3002\u3053\u308c\u306f\u3001<strong>LLM<\/strong>\u304c\u8aa4\u60c5\u5831\u3001\u30d8\u30a4\u30c8\u30b9\u30d4\u30fc\u30c1\u3001\u307e\u305f\u306f\u305d\u306e\u4ed6\u306e\u671b\u307e\u3057\u304f\u306a\u3044\u51fa\u529b\u3092\u751f\u6210\u3059\u308b\u30ea\u30b9\u30af\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306e\u91cd\u8981\u306a\u30b9\u30c6\u30c3\u30d7\u3067\u3059\u3002<\/p>\n<p>\u3053\u306e\u30a2\u30d7\u30ed\u30fc\u30c1\u306f\u3001\u30b5\u30a4\u30d0\u30fc\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u306b\u304a\u3051\u308b\u300c\u30ec\u30c3\u30c9\u30c1\u30fc\u30e0\u300d\u306e\u6982\u5ff5\u306b\u4f3c\u3066\u304a\u308a\u3001\u653b\u6483\u8005\u306e\u8996\u70b9\u304b\u3089\u30b7\u30b9\u30c6\u30e0\u3092\u30c6\u30b9\u30c8\u3059\u308b\u3053\u3068\u3067\u9632\u5fa1\u3092\u5f37\u5316\u3057\u307e\u3059\u3002<strong>GPT-Red<\/strong>\u306f\u3001\u4eba\u9593\u306e\u30ec\u30c3\u30c9\u30c1\u30fc\u30e0\u304c\u767a\u898b\u3059\u308b\u306e\u306b\u6642\u9593\u304c\u304b\u304b\u308b\u3001\u307e\u305f\u306f\u898b\u843d\u3068\u3059\u53ef\u80fd\u6027\u306e\u3042\u308b\u8907\u96d1\u306a\u653b\u6483\u30d9\u30af\u30c8\u30eb\u3092\u81ea\u52d5\u7684\u306b\u751f\u6210\u3057\u3001\u30c6\u30b9\u30c8\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n<p><strong>OpenAI<\/strong>\u306f\u3001\u3053\u306e\u3088\u3046\u306a\u5185\u90e8\u7684\u306a\u300c\u30cf\u30c3\u30ad\u30f3\u30b0\u300d\u30c4\u30fc\u30eb\u3092\u4f7f\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u30e2\u30c7\u30eb\u304c\u4e00\u822c\u306b\u516c\u958b\u3055\u308c\u308b\u524d\u306b\u6f5c\u5728\u7684\u306a\u60aa\u7528\u7d4c\u8def\u3092\u7279\u5b9a\u3057\u3001\u4fee\u6b63\u3059\u308b\u3053\u3068\u3092\u76ee\u6307\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u3088\u308a\u5b89\u5168\u3067\u4fe1\u983c\u6027\u306e\u9ad8\u3044<strong>AI<\/strong>\u30b7\u30b9\u30c6\u30e0\u306e\u958b\u767a\u304c\u4fc3\u9032\u3055\u308c\u307e\u3059\u3002<\/p>\n<p><strong>GPT-Red<\/strong>\u306e\u3088\u3046\u306a\u30c4\u30fc\u30eb\u306e\u958b\u767a\u306f\u3001<strong>AI<\/strong>\u306e\u80fd\u529b\u304c\u9032\u5316\u3059\u308b\u306b\u3064\u308c\u3066\u3001\u305d\u306e\u5b89\u5168\u6027\u3068\u8cac\u4efb\u3042\u308b\u5c55\u958b\u3092\u78ba\u4fdd\u3059\u308b\u305f\u3081\u306e\u7d99\u7d9a\u7684\u306a\u53d6\u308a\u7d44\u307f\u306e\u91cd\u8981\u6027\u3092\u6d6e\u304d\u5f6b\u308a\u306b\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306f\u3001<strong>LLM<\/strong>\u304c\u793e\u4f1a\u306b\u4e0e\u3048\u308b\u5f71\u97ff\u3092\u8003\u616e\u3057\u305f\u3001\u30d7\u30ed\u30a2\u30af\u30c6\u30a3\u30d6\u306a\u5b89\u5168\u5bfe\u7b56\u306e\u4e00\u4f8b\u3067\u3059\u3002<\/p>\n<p>\ud83d\udc49 <strong><a href=\"https:\/\/www.technologyreview.com\/2026\/07\/15\/1140514\/meet-gpt-red-an-llm-super-hacker-openai-built-to-make-its-models-safer\/\" target=\"_blank\" rel=\"noopener\">Technology Review \u3067\u8a18\u4e8b\u5168\u6587\u3092\u8aad\u3080<\/a><\/strong><\/p>\n<\/div>\n<ul>\n<li><strong>\u8981\u70b9:<\/strong> OpenAI's GPT-Red is an LLM designed to act as a 'super-hacker,' proactively identifying vulnerabilities and 'jailbreak' prompts in other LLMs to enhance their safety and prevent the generation of harmful content.<\/li>\n<li><strong>\u8457\u8005:<\/strong> Will Douglas Heaven<\/li>\n<\/ul>\n<blockquote class=\"wp-block-quote\"><p><span>English Summary:<\/span><\/p>\n<p><strong>OpenAI<\/strong> has developed a 'super-hacker' <strong>LLM<\/strong> named <strong>GPT-Red<\/strong>, designed to enhance the safety of large language models (<strong>LLMs<\/strong>). This system is engineered to identify vulnerabilities and 'jailbreak' prompts that malicious users might exploit to bypass <strong>LLM<\/strong> guardrails and generate harmful content.<\/p>\n<p>By actively seeking out weaknesses in existing <strong>LLMs<\/strong>, <strong>GPT-Red<\/strong> helps <strong>OpenAI<\/strong> strengthen the models' robustness and ensure adherence to ethical guidelines. This represents a crucial step in mitigating the risks of <strong>LLMs<\/strong> producing misinformation, hate speech, or other undesirable outputs.<\/p>\n<p>This approach is akin to the 'red teaming' concept in cybersecurity, where systems are tested from an attacker's perspective to bolster defenses. <strong>GPT-Red<\/strong> can automatically generate and test complex attack vectors that human red teams might take longer to discover or potentially overlook.<\/p>\n<p>Through the use of such internal 'hacking' tools, <strong>OpenAI<\/strong> aims to identify and rectify potential misuse pathways before models are released to the public. This fosters the development of safer and more reliable <strong>AI<\/strong> systems.<\/p>\n<p>The development of tools like <strong>GPT-Red<\/strong> underscores the ongoing importance of ensuring <strong>AI<\/strong> safety and responsible deployment as <strong>AI<\/strong> capabilities evolve. It serves as an example of proactive safety measures taken in consideration of the societal impact of <strong>LLMs<\/strong>.<\/p>\n<\/blockquote>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>AI\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u306b\u3088\u308bAI\u8a13\u7df4\u306e\u81ea\u52d5\u5316\u3001\u91cf\u5b50\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u306b\u3088\u308b\u6838\u878d\u5408\u7814\u7a76\u306e\u52a0\u901f\u3001\u305d\u3057\u3066OpenAI\u306eGPT-Red\u306b\u3088\u308bLLM\u5b89\u5168\u6027\u306e\u5411\u4e0a\u306b\u95a2\u3059\u308b\u6700\u65b0\u6280\u8853\u52d5\u5411\u3092\u6df1\u6398\u308a\u3057\u307e\u3059\u3002<\/p>\n","protected":false},"author":1,"featured_media":855,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"sns_share_botton_hide":"","vkExUnit_sns_title":"","_vk_print_noindex":"","sitemap_hide":"","_veu_custom_css":"","veu_display_promotion_alert":"","vkexunit_cta_each_option":"","_lightning_design_setting":[],"footnotes":""},"categories":[3],"tags":[8,17,58,16,57,110,26,225,15,205],"class_list":["post-1093","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-columns","tag-ai","tag-anthropic","tag-llm","tag-openai","tag-110","tag-26","tag-225","tag-205"],"veu_head_title_object":{"title":"","add_site_title":""},"_links":{"self":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/posts\/1093","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/comments?post=1093"}],"version-history":[{"count":0,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/posts\/1093\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/media\/855"}],"wp:attachment":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/media?parent=1093"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/categories?post=1093"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/tags?post=1093"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}