{"id":9649,"date":"2025-01-27T20:29:33","date_gmt":"2025-01-28T01:29:33","guid":{"rendered":"https:\/\/www.inthacity.com\/blog\/uncategorized\/bytedance-ai-dominates-openai-operator-ui-tars-hands-free\/"},"modified":"2025-08-23T19:53:01","modified_gmt":"2025-08-24T00:53:01","slug":"bytedance-ai-dominates-openai-operator-ui-tars-hands-free","status":"publish","type":"post","link":"https:\/\/www.inthacity.com\/blog\/tech\/ai\/bytedance-ai-dominates-openai-operator-ui-tars-hands-free\/","title":{"rendered":"ByteDance\u2019s New AI Dominates OpenAI Operator – UI-TARS – Hands-Free AI"},"content":{"rendered":"

If you\u2019ve ever dreamed of having a personal assistant that can handle everything from booking flights to editing Photoshop files, your dreams just got a lot closer to reality. Thanks to AI Revolution\u2019s latest video, we\u2019re diving into UI Tar<\/strong>, an AI agent that doesn\u2019t just talk\u2014it does<\/em>. This groundbreaking technology is here to revolutionize the way we interact with our devices, and it\u2019s more impressive (and slightly terrifying) than you might imagine.<\/p>\n

Developed by ByteDance<\/a> in collaboration with Tsinghua University<\/a>, UI Tar is a native GUI agent that can literally take control of your computer or phone. Whether you\u2019re on a Mac or PC, this AI can navigate interfaces, perform complex tasks, and even fix its own mistakes. It\u2019s like having a super-smart coworker who never sleeps, never complains, and\u2014most importantly\u2014never messes up your coffee order.<\/p>\n

What Makes UI Tar So Special?<\/h2>\n

UI Tar is available in two versions: one with 7 billion parameters and another with a whopping 72 billion parameters. Trained on a massive dataset of about 50 billion tokens, this AI isn\u2019t just generating text\u2014it\u2019s controlling your graphical user interface (GUI) in real-time. Imagine saying, \u201cFind me roundtrip flights from Seattle to New York next month,\u201d and watching as UI Tar opens Delta\u2019s website<\/a>, fills out the details, selects the dates, filters by price, and even clicks around the site as needed. It\u2019s not just efficient\u2014it\u2019s downright magical.<\/p>\n

What sets UI Tar apart is its \u201cpure vision-based agent\u201d method. Unlike older AI systems that rely on text-based data like HTML or accessibility trees, UI Tar perceives the screen visually, just like a human. It doesn\u2019t need to peek behind the curtain of code\u2014it sees a screenshot, understands the layout, and interacts with it as if it\u2019s a real user. This makes it incredibly flexible, allowing it to adapt to changes in the interface or platform without missing a beat.<\/p>\n

How Does UI Tar Outperform the Competition?<\/h3>\n

When it comes to benchmarks, UI Tar is a serious contender. It outperforms giants like OpenAI\u2019s GPT-4<\/a>, Anthropic\u2019s Claude<\/a>, and even Google\u2019s Gemini<\/a> on more than 10 different GUI benchmarks. For example, on the Visual Web Bench, UI Tar scored 82.8 compared to GPT-4\u2019s 78.5. It also shines on multi-step tasks, like rearranging slides in PowerPoint or customizing mobile app settings, where it scored 24.6 on the OS World benchmark\u2014significantly higher than Claude\u2019s 22.0.<\/p>\n

Part of its success lies in \u201creflection tuning,\u201d a process where the AI corrects its own errors. If it tries to install a Visual Studio Code<\/a> extension and something goes wrong, UI Tar notices the glitch, checks if the app is still loading, and adjusts its approach. This iterative feedback process means the AI gets better with every mistake, polishing its performance like a seasoned pro.<\/p>\n\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t
<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t
<\/div>\n\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\n\t\t\t\t\t
\nPrev<\/a>\n1<\/span> of<\/span> 1<\/span>\nNext<\/a>\n<\/div>\n
\n\t