Claude Sonnet 5以智能体（Agent，指能自主规划并执行多步操作的AI系统）身份，从候选基因疗法中自主选择了药物、目标脑区和注射坐标，并对一只癫痫模型小鼠完成了手术。Exin Therapeutics联合创始人、CEO兼牛津大学博士生Gabriel Ocana-Santero描述了这一过程。据the-decoder报道，这是AI智能体首次自主选定治疗方案并将其施用于活体哺乳动物。
手术过程中出现了意外。连接一度中断，智能体重新计算后发现目标位置超出了设备的物理活动范围，随即拒绝继续操作。Ocana-Santero未能说服它继续执行，尽管麻醉状态的小鼠仍躺在手术台上，他最终只能手动修复代码，解决问题。
安全设置上，持证兽医批准了每一个不可逆步骤，并手持物理紧急停止装置，一旦触发会立即切断所有电机的电源。据Ocana-Santero介绍，小鼠术后恢复良好。
模型选择上的差异
Ocana-Santero测试了多个Claude模型执行这台实时手术，Sonnet 5是唯一愿意从头到尾完成操作的模型，能力更强的Opus 5.5和Opus 5都拒绝执行。aihot.news的报道没有说明两款Opus模型拒绝的具体理由，也没有给出风险判断机制上的技术解释。这种差异可能和不同模型对不可逆操作的风险评估阈值不同有关，但这属于推测，源文本没有提供支撑，需要Exin团队公开更多训练或测试细节才能确认。
尚未解决的问题
这项结果尚未经过同行评审。智能体自主选择治疗方案的决策过程、出错后修复代码的具体内容、以及不同模型在风险判断上的差异，目前的信息来自Ocana-Santero本人的一次描述，aihot和the-decoder是独立转述这段描述的两个报道渠道，还没有看到第三方研究者或业内同行对这套智能体手术流程的技术评价。这类系统对其他实验室是否可复现、紧急停止装置之外还有哪些安全约束，源文本均未提及，仍是空白。
原文地址：tools.wflynn.cn/aitools/tre…
关注公众号「YourTools」，每天推送 GitHub 趋势与开源工具速报：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Claude Sonnet 5 as AI agent (Agent, an AI system capable of autonomous planning and implementing multi-step operations) has independently selected drugs, target brain areas and injection coordinates from candidate gene therapy and has performed surgery on an epilepsy model mouse. The process was described by Gabriel Ocana-Santero, co-founder of Exin Therapeutics, CEO and doctoral student at Oxford University. According to the-decoder, this was the first time that AZAI agent had chosen its own treatment and applied it to live mammals.
An accident occurred during the operation. Once the connection was interrupted, AI agent recalculated and found that the target position was beyond the physical activity of the equipment and then refused to proceed. Ocana-Santero couldn't convince it to continue, even though the anaesthetized mice were still lying. On the operating table, he eventually had to repair the code manually and solve the problem.
On the safety set-up, the certified veterinarian approved every irreversible step and armed with a physical emergency halter, which, if triggered, would cut off all electrical power. According to Ocana-Santero, the mouse recovered well after the operation.
Difference in Model Selection
Ocana-Santero tested several Claude models for this real-time operation, and Sonnet 5 is the only model willing to complete the operation from the beginning to the end. The more capable Opus 5.5 and Opus 5 refuse to perform. The report of aihot.news does not provide specific reasons for the rejection of the two Opus models, nor does it provide a technical explanation for the risk judgement mechanism. This difference may be related to different risk assessment thresholds for irreversible operations from different models, but it is speculative that the source text is not supported and requires additional training or test details to be disclosed by the Exin team for confirmation.
Outstanding issues
This result has not yet been peer-reviewed. The decision-making process of AI agent's autonomous choice of treatment options, the specific content of the error repair code, and the differences in risk judgement between different models are currently derived from a description by Ocana-Santero himself, with aihot and the-decoder being the two reporting channels that independently reproduce this description, and no technical evaluation of the AI agent procedure by third-party researchers or industry counterparts has been seen. The source text does not mention any safety constraints other than the re-emergence of other laboratories and the emergency cessation mechanism.
Original address:tools.wflynn.cn/aitools/tre...
Following the WeChat Official Account "YourTools", GitHub's trend and open source tool is sent daily:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.