上海交大等发布科研大模型 BigBang-V1，让 AI 参与造题并自我迭代
BigBang-V1 用 AI 合成数据并自动迭代训练任务，证明了小模型可通过自我造题逼近乃至超越更大模型。
上海交大等机构发布科研大模型 BigBang-V1，后训练数据完全由 AI 合成。该系统让 AI 扮演“出题者”和“考官”，自动生成、筛选并迭代训练任务，用科学问题的可验证性驱动模型自我进化。在 35B 规模模型中，BigBang-V1 于 10 项评测取得最高分，部分任务超更大模型。
正文摘录
35B Model Outperforms Trillion-Parameter AI — Shanghai Jiao Tong AI Starts Creating Its Own Problems and Iterating Editor | +0, Yang Wen — AI has begun participating in the creation of AI, and the next question is what it should learn next. Ilya Sutskever pointed out late last year that pretraining data will eventually be exhausted. Simply scaling existing training runs by a factor of 100 is unlikely to yield fundamental change. After years of scale expansion, the AI industry is re-entering a "research era," where the next capability leap will require new training methods and learning mechanisms. [](https://mp.weixin.qq.com/s/w82zfOzmAL4QePf9hQWhHA) Video source: [https://www.youtube.com/…