{"id":78038,"date":"2026-03-17T03:28:00","date_gmt":"2026-03-17T07:28:00","guid":{"rendered":"https:\/\/blog.cyberconservices.com\/?p=78038"},"modified":"2026-03-15T16:35:29","modified_gmt":"2026-03-15T20:35:29","slug":"the-ai-that-taught-itself","status":"publish","type":"post","link":"https:\/\/blog.cyberconservices.com\/index.php\/2026\/03\/17\/the-ai-that-taught-itself\/","title":{"rendered":"The AI That Taught Itself"},"content":{"rendered":"

By Magali Gruet – A\u00a0<\/span>new study<\/span><\/a>\u00a0from the\u00a0<\/span>USC Viterbi School of Engineering<\/span><\/a>\u00a0was accepted at the\u00a0IEEE SoutheastCon 2026<\/a>, taking place March 12-15. It\u00a0<\/span>suggests something far more surprising: with the right method in place, an AI model can dramatically improve its performance in territory it was barely trained on, pushing well past what its training data alone would ever allow.<\/span><\/p>\n

The method was developed by\u00a0<\/span>Minda Li<\/span><\/a>, a USC Viterbi undergraduate who has been pursuing research since her freshman year, working alongside her advisor\u00a0<\/span>Bhaskar Krishnamachari<\/span><\/a>*, a Faculty Fellow and Systems Professor in the\u00a0Ming Hsieh Department of Electrical and Computer Engineering<\/a>, with a joint appointment in the\u00a0Thomas Lord Department of Computer Science<\/a>\u00a0at the\u00a0USC Viterbi School of Engineering<\/a>\u00a0and the\u00a0USC School of Advanced Computing<\/a>. Together, they tested GPT-5\u2019s ability to write code in Idris, an extraordinarily obscure programming language with a fraction of the online presence of mainstream languages like Python. The results were striking: by giving the AI feedback on its errors and letting it try again, Li pushed the model\u2019s success rate from a dismal 39% all the way to 96%.<\/span><\/p>\n

A Language So Obscure, Even the Researchers Didn\u2019t Know It<\/h2>\n

Python, the world\u2019s most popular programming language, has over 24 million code repositories publicly available online, a vast library that AI models like GPT-5 learn from during training. Idris, the language Li and Krishnamachari chose to test, has approximately 2,000. That is roughly 10,000 times less data.<\/p>\n

The choice of Idris was deliberate, and, as Krishnamachari describes it, a little playful. \u201cWe were hunting for a language so obscure that we hadn\u2019t heard of it,\u201d he said. \u201cI think we were just in my office together, googling around, trying to find some crazy language that no one\u2019s ever heard of.\u201d They found Idris, a dependently typed functional programming language used by a small community of specialists, and decided it was the perfect test case.<\/p>\n

Crucially, neither researcher could write a line of it themselves. \u201cNeither Minda nor I had ever coded in it, and frankly, we could not tell you if the code was correct or wrong,\u201d Krishnamachari admitted. That is part of what makes the findings so striking: Li was guiding an AI to master a language that its own guides could not speak.<\/p>\n

The Breakthrough: A Feedback Loop That Changes Everything<\/h2>\n

Li started by simply asking GPT-5 to solve 56 Idris coding exercises on Exercism, a popular coding practice platform. Out of the box, the model solved only 22 of them, a 39% success rate, far below its 90% success rate in Python and 74% in Erlang.<\/p>\n

She then tried several approaches to improve performance: providing documentation, error manuals, and reference guides. These helped somewhat, pushing the success rate to the low 60s, but never dramatically.<\/p>\n

The breakthrough came when she implemented what they call a compiler feedback loop. A compiler is the software that translates human-written code into instructions a computer can execute. When code is wrong, the compiler says so, precisely and in technical detail. She began capturing those error messages and feeding them directly back to GPT-5, asking it to fix the specific problems identified and try again. Up to 20 times per problem.<\/p>\n

\u201cI thought we\u2019d probably get a 10% jump,\u201d said Li, who designed and ran the experiments. \u201cI was surprised that just that alone, seemingly one simple thing, just keep recompiling, keep trying, was able to get to 96%.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"

For years, the guiding assumption of artificial intelligence has been simple: an AI is only as good as the data it has seen. Feed it more, train it longer, and it performs better. Feed it less, and it stumbles.\u00a0<\/p>\n Continue reading →<\/span><\/a>","protected":false},"author":1,"featured_media":78039,"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,"advanced_seo_description":"","jetpack_seo_html_title":"","jetpack_seo_noindex":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[638],"tags":[],"class_list":["post-78038","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"https:\/\/blog.cyberconservices.com\/wp-content\/uploads\/2026\/03\/photographic-quality-full-color-image-of.png","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/posts\/78038","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/comments?post=78038"}],"version-history":[{"count":2,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/posts\/78038\/revisions"}],"predecessor-version":[{"id":78041,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/posts\/78038\/revisions\/78041"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/media\/78039"}],"wp:attachment":[{"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/media?parent=78038"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/categories?post=78038"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.cyberconservices.com\/index.php\/wp-json\/wp\/v2\/tags?post=78038"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}