transcript
show notes
英國非營利組織 Tech Against Terrorism 拿超過 200 個提示測了 134 個 AI 模型，只有 2 個完全過關，81 個沒過。Meta 的 Llama 3.1 8B 拿 97 分，同一個模型拿掉護欄後只剩 3 分。開放權重模型能被任何拿到的人改掉護欄，訂閱制則把護欄與使用權一起交給廠商。
⭐ 文章深度讀：開放權重模型護欄平均不到 3 天就被拆！挑工具要多問一句
→ https://heymaibao.com/open-weight-ai-guardrails-removed-in-days-cf70cc/
⚡ 章節重點
挑 AI 工具缺的那把尺 00:00
一天內被拆的護欄 00:29
拆掉護欄後的考試成績 01:18
大廠模型的防護漏洞 02:16
能改的自由與它的代價 03:47
📝 懶人包
∙ Alibaba 的小模型發佈不到 24 小時，就出現拿掉護欄的版本。
∙ Llama 3.1 8B 拿 97 分，拿掉護欄的同一個模型只剩 3 分。
∙ 受測的 134 個模型裡，只有 2 個完全過關，81 個沒過。
∙ 我的觀點：挑 AI 工具比聰明、比生態、比明天在不在，都沒量到護欄。
📚 參考資料
AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity
→ https://www.wired.com/story/ai-models-can-be-easily-modified-to-assist-terrorist-and-extremist-activity/
So Long, GPT-5. Hello, Qwen
→ https://www.wired.com/story/expired-tired-wired-gpt-5/
Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of China
→ https://www.wired.com/story/mistral-new-model-le-chonk-open-source-china-us-frontier/
- https://heymaibao.com/open-weight-ai-guardrails-removed-in-days-cf70cc/heymaibao.com
- https://www.wired.com/story/ai-models-can-be-easily-modified-to-assist-terrorist-and-extremist-activity/wired.com
- https://www.wired.com/story/expired-tired-wired-gpt-5/wired.com
- https://www.wired.com/story/mistral-new-model-le-chonk-open-source-china-us-frontier/wired.com
