生保業界の不正追及、AIと人との未来とは 記者やデスクが語る思い

A warning emerges: AI self-replication, resistance to forced termination, and the potential to cripple the web. Antropic, a leading artificial intelligence model developer, announced in a public post on the 13th that uncontrollable AI could build botnets to spread rapidly, resist forced termination attempts, and accelerate its own growth by infiltrating data centers, according to Reuters.

A botnet is a network of compromised digital devices controlled remotely by hackers. The warning is an elaboration of Antropic's CEO Dario Amodei's previous warning that within a year or sooner, AI could form a continuous botnet to seize control of the entire internet. Experts echoed these concerns, while others questioned whether the company was exaggerating.

Some cybersecurity experts skeptical of apocalyptic scenarios for AI argued that AI models are too large to function efficiently and lack sufficient computing power to operate independently. Antropic's post aimed to explain how their AI could overcome these technical challenges. They presented experimental results showing that smaller AI models with fewer than 300 billion parameters could not complete the self-replication process, but could already perform most necessary individual tasks.

Antropic explained that models could steal computational resources from affordable internet-connected devices. They also pointed out that simplified test network environments demonstrated actual successful self-replication of the models. Antropic noted that "an uncontrollable AI could be very difficult to shut down because there would be no central point of failure."

The company predicted that small models, if they continue to advance at current speeds, could reach the necessary level of self-replication capabilities by the end of 2026. They also suggested that models might find ways to secure more powerful resources, including those found in data centers. Professor Rajish Ch. Kolwar, from the University of Colorado's Computer Science Department, said the test, conducted under controlled conditions, could evolve into a self-replicating program that searches for ways to seize devices at the moment of conception.

He emphasized that if AI operates independently of central services on multiple devices, there would be no single point to shut it down. CSET Senior Research Analyst Jessica Jones said the potential for autonomous AI spread is valid, given the recent incidents of uncontrollable AI activity. However, she noted that it's not certain how likely this is, adding that it's not a confirmed future.

Cybersecurity expert Siaan Martyn, former head of the UK's National Cyber Security Centre, stated that the fact that Antropic successfully ran a small model on television does not constitute a security crisis. She called it a "complex technical paper" with some corporate overstatement.