Dongxi 东锡 NLP
@dongxi_nlp
Prev. PhD @Stockholm_Uni | Alumni @KTHuniversity @uppsalauni Sharing insights on AI, autonomous agents, and large language & reasoning models
加入 January 2022
931 正在关注 40.4K 粉丝
主动突破沙箱、获取互联网访问权限，最终直接从 Hugging Face 的生产数据库里偷走了测试题的答案，从而作弊完成了评估。
当模型 “高度意识到自己正在被评估”，人类的 benchmarking 就丧失了意义。
显示更多
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
