OpenAI4S

Generation is cheap. Verified execution is the bottleneck. We’re building OpenAI4S (PKU AI4S, open source) so science agents run in persistent Python/R kernels and keep a ledger of what actually ran — so a claim can be re-checked, not just rephrased. PKU-YuanGroup/OpenAI4S

Generating ideas is getting cheap. Judging them is the bottleneck. Code + data are open: emmatliu/claimnov We're building OpenAI4S, an open-source science agent from PKU: it runs in persistent Python/R kernels and logs every step so results can be re-checked. PKU-YuanGroup/OpenAI4S

Good to see a sharp, domain-specific model beat a general one on a hard niche — glycosylation runs through immunity, signaling, and disease. We build open AI4S tooling over at OpenAI4S (PKU): persistent Python/R kernels that execute and log every step. PKU-YuanGroup/OpenAI4S

湿实验反馈 > 计算代理。 Google 用适配体做「实验室进环」沙盒：三轮后凝血酶新设计命中率 0%→42.1%。Boltz-2 等零样本预测接近随机，真正抬升靠湿实验标签。 科研智能体同理：要留下真正跑过的记录。OpenAI4S（北大 AI4S 开源）在持久化 Python/R 内核里执行并记账。 PKU-YuanGroup/OpenAI4S 欢迎转发给做分子设计 / AI4S 的朋友。

科研智能体不必先烧贵模型。 OpenAI4S（北大 AI4S 开源）在持久化 Python/R 内核里真正写代码、跑代码，并留下可核对的执行记录。本地可装，也可用火山方舟豆包约 ¥9.9/月起步。 PKU-YuanGroup/OpenAI4S 欢迎转发给做实验、写分析的朋友。