计算机科学 on 问衿的学习笔记https://wenjinji.cn/en/subjects/%E8%AE%A1%E7%AE%97%E6%9C%BA%E7%A7%91%E5%AD%A6/Recent content in 计算机科学 on 问衿的学习笔记HugoenFri, 07 Aug 2026 00:00:00 +0000I Told an AI It Could Only See Pixels and Click. Then I Watched It Get Useful. — Qwen-CUA Stress Testhttps://wenjinji.cn/en/episodes/qwen-cua/Fri, 07 Aug 2026 00:00:00 +0000https://wenjinji.cn/en/episodes/qwen-cua/Given only a screen, mouse and keyboard, this AI intern still books flights and fills forms — its OSWorld score jumped from 38 to 86.2.AI Video Is Indistinguishable From Real Footage, But Does It Actually Know the Culture? A 12-Country, 1,000-Prompt Stress Testhttps://wenjinji.cn/en/episodes/culturevid/Thu, 06 Aug 2026 00:00:00 +0000https://wenjinji.cn/en/episodes/culturevid/A 12-country, 1000-prompt benchmark on 7 text-to-video models: the footage looks real, but the culture is often wrong.Your AI Phone Never Stops Listening—and Someone Is Squeezing Commands Into the Airhttps://wenjinji.cn/en/episodes/audio-inject/Thu, 06 Aug 2026 00:00:00 +0000https://wenjinji.cn/en/episodes/audio-inject/2,160 inaudible attack clips tested on 11 top AI agents: the smarter the model, the easier to fool — the best one fell for 7 out of 10.Who Audits the AI Therapist? A Clinical Audit of Nine Chatbotshttps://wenjinji.cn/en/episodes/sim-vail/Tue, 04 Aug 2026 00:00:00 +0000https://wenjinji.cn/en/episodes/sim-vail/810 conversations, 6,000+ turns: nine chatbots got their first clinical checkup as ‘AI therapists’ — the AI grader agreed with 27 psychiatrists more than they agreed with each other.Skill Self-Play: When LLMs Learn to Quiz Themselves — How Three Roles Co-Evolvehttps://wenjinji.cn/en/episodes/skill-self-play/Tue, 28 Jul 2026 00:00:00 +0000https://wenjinji.cn/en/episodes/skill-self-play/Alibaba’s Skill Self-Play lets LLMs write and grade their own exams: a weak base model soared from 20.7 to 63.6 (+42.9), while a strong one gained just 2.8 points.