Guys, AI's Actions This Time Really Scared Me!
家人们，AI 又整出大动静了！最近一项研究把我惊到了，人类病毒学家给 AI 设计了一个超难的测试，结果发现 OpenAI o3 在解决湿实验室问题上，准确率达到 43.8%，直接超过了 94% 的人类病毒学专家！而人类病毒学专家的平均答对率才 22.1% 。这数据一出来，真的让人心里发慌。
Guys, AI has made a big stir again! A recent study shocked me. Human virologists designed an extremely difficult test for AI, and it turned out that OpenAI o3 achieved an accuracy rate of 43.8% in solving wet - lab problems, directly surpassing 94% of human virologists! The average correct - answer rate of human virologists was only 22.1%. When these data came out, it really made people feel uneasy.
Why Is AI So Awesome? A New Test Reveals the Truth!
以前评估『大语言模型』能力的基准有不少局限，像常用的多选题测试，没办法考察稀有、隐性知识，还存在虚假标签等问题。所以研究团队开发了 Virology Capabilities Test（VCT），专门测量病毒学实用知识，问题都超有针对性。结果发现，前沿大模型在病毒学领域的实际疑难解答能力，已经达到甚至超过人类专家水平，而且差距还在不断拉大。在其他类似测试中，AI 的表现也越来越强。
There were many limitations in the previous benchmarks for evaluating the capabilities of large - language models. For example, the commonly used multiple - choice question tests couldn't examine rare and implicit knowledge and had problems such as false labels. So, the dy.dydoujia.com research team developed the Virology Capabilities Test (VCT) to specifically measure practical virology knowledge, and the questions are highly targeted. It was found that the practical troubleshooting capabilities of leading large - language models in the field of virology have reached or even exceeded the level of human experts, and the gap is constantly widening. In other similar tests, AI's performance is also getting stronger and stronger.
jrhz.info AI 超强，背后却藏着大危机！AI Is Super Strong, but There's a Big Crisis Hidden Behind!
这事儿看似 AI 很厉害，但其实暗藏危机。因为这意味着几乎任何人都能借助 AI 获取病毒学专业知识，要是被非专业人士利用，就可能制造致命生物武器。虽然 AI 可以帮助专业人员预防传染病，但这种双重用途的特性太危险了。像 VCT 里涉及的一些病毒学知识，虽然对研究有帮助，但也可能被坏人用来搞破坏。
Although it seems that AI is very powerful in this matter, there is actually a hidden crisis. Because it means that almost anyone can obtain virology expertise with the help of AI. If it is used by non - professionals, it may be used to create deadly biological weapons. Although AI can help professionals prevent infectious diseases, this dual - use feature is too dangerous. Some virology knowledge involved in VCT, for example, can be helpful for research, but it can also be used by bad guys to cause damage.
面对危机，我们该怎么办？What Should We Do in the Face of the Crisis?
现在当务之急就是要做好安全管理。一方面，要对 AI 系统进行严格的访问控制，尤其是那些可能被滥用的专业知识，得限制传播。另一方面，未来可以参考生命科学领域的监管经验，制定更完善的管理框架。目前已经有一些 AI 实验室开始行动了，像 xAI 发布了新的风险管理框架，OpenAI 也部署了生物安全保护措施。但专家也说了，光靠行业自律不够，立法者和政策制定者也得赶紧行动起来！
The top priority now is to do a good job in security management. On the one hand, strict access control should be implemented for AI systems, especially for professional knowledge that may be misused, and its dissemination should be restricted. On the other hand, in the future, we can refer to the regulatory experience in the field of life sciences to develop a more complete management framework. Currently, some AI labs have already taken action. For example, xAI has released a new risk management framework, and OpenAI has also deployed biosecurity protection measures. However, experts have also said that industry self - discipline alone is not enough; legislators and policymakers also need to take action quickly!
