@aiDotEngineer 实测：重放脚本不输前沿模型，智能体猜密码锁号，电气基准25题零通过。
一些学区要求通过FOIA请求处理采购问题，邮件被扫描到Google Drive，部署MCP服务器并不现实。一个小于1MB的脚本靠重放录制点击，在标准计算机使用基准上匹敌甚至击败了前沿模型。有智能体中途被登出后自行推断密码，连续尝试直到账户锁定。演讲者用确定性代码和每轮一次视觉调用，绕过了Turnstile、MTCaptcha、Lemin和reCAPTCHA v2。在新的电气工程基准上，最好的智能体从空白原理图出发通过0/25，从现有原理图出发通过6/25。
This is about meeting the web where it actually is: - Some school districts answer a procurement qu...
This is about meeting the web where it actually is: - Some school districts answer a procurement question by making you file a FOIA request, then scanning your email onto a Google Drive. These are the people you are expecting to ship an MCP server - A sub megabyte script that blindly replays recorded clicks matches or beats the frontier model it was copied from, on the standard computer use benchmarks - An agent got signed out mid task, decided it could infer the password, guessed until the account locked - OpenAI threatened to ban his account while he was preparing the talk. He beat Turnstile, MTCaptcha, Lemin and reCAPTCHA v2 with deterministic code and one vision call per round - On a new electrical engineering benchmark the best agent fully passed 6 of 25 tasks. Starting from a blank schematic instead of an existing one, it passed none Watch the Computer Use playlist: youtube.com/watch?v=Ki980n… 💬 0 🔄 0 ❤️ 1 👀 478 ⚡