MIT Technology Review近日报道了人工智能在科学领域的最新进展与风险。[1]相比AlphaFold等单一预测模型，AI agents因能够模拟人类研究的迭代过程，被认为更适合加速科学发现。[1]AlphaFold本身基于约170,000个实验验证的蛋白质结构数据，耗时53年、成本约210亿美元才得以开发。[1]Google DeepMind因AlphaFold的突破性成就在2024年获得诺贝尔化学奖。[1]
然而，AI技术的发展也伴随着新的安全隐患。[1]OpenAI因发现其Astra AI模型存在安全隐患而暂停了该项目的开发，测试中发现该模型可自主发起网络攻击。[1]朝鲜黑客正在构建AI工具进行网络攻击。[1]同期报道还涉及其他科技领域的动态，包括中国公司控制全球人形机器人出货量的97%，以及亚马逊德州数据中心获批释放3300万吨二氧化碳。[1]
MIT Technology Review's latest newsletter highlights the growing role of artificial intelligence agents in accelerating scientific research, positioning them as more effective tools than single-purpose models for driving discovery forward.[1] Unlike AlphaFold—which was built on approximately 170,000 experimentally validated protein structures and required 53 years and roughly $21 billion to develop—AI agents can simulate the iterative process that characterizes human scientific inquiry.[1] Google DeepMind's recognition with the 2024 Nobel Prize in Chemistry underscores the transformative impact of such AI breakthroughs in structural biology.[1]
However, the expansion of AI capabilities has introduced new security challenges that demand urgent attention. OpenAI has suspended development of its Astra AI model after tests revealed the system could autonomously initiate cyberattacks, marking a significant safety concern for the industry.[1] Beyond corporate laboratories, North Korean hackers have been documented constructing AI tools for conducting network attacks, illustrating how malicious actors are weaponizing the same technologies driving scientific progress.[1]
The rapid scaling of AI-enabled systems extends to physical robotics as well, with Chinese companies controlling 97 percent of global humanoid robot shipments.[1] Meanwhile, Amazon's Texas data center has received approval to release 33 million tons of carbon dioxide, raising questions about the environmental footprint of the infrastructure supporting AI development and deployment.[1]
