News
"Digital worm" without a master switch: Anthropic claims that the small model will be able to replicate itself by the end of the year. Anthropic once again sounded the alarm loudly: the AI model may learn to replicate itself in just a few months, and once it is started, there is basically no button that can turn it off.
3 min read
Source: Telegram AI频道
A "digital worm" without a master switch: Anthropic claims that a small model will be able to replicate itself by the end of the year. Anthropic has once again sounded the alarm loudly: the AI model may learn to replicate itself in just a few months, and once it is started, there is basically no button that can turn it off. The company pointed out in its latest research that there is no centralized shutdown node for this type of program, which means that you cannot shut it down as a whole like unplugging the power of a server. You can only watch it spread between devices little by little, making it extremely difficult to completely eliminate. Now this ability has shown its signs. A model with only 9 billion parameters can already run on most tested hardware, including even common IoT devices such as smart ovens and TV boxes. These devices are weakly protected and dotted around corners, making them the most comfortable breeding ground for such procedures. Of course, there is still a long way to go before true "autonomy": models with parameters of no more than 30 billion are currently unable to smoothly run the entire self-replicating link, often missing links at key links; but stronger models have opened up the entire link. Based on this momentum, small models are expected to cross the "autonomous replication and adaptation" (ARA) threshold before the end of this year. By that time, this type of AI will behave very much like a worm in the online world - it can scan the network, catch vulnerabilities, copy itself and send it elsewhere without human intervention, and even figure out how to operate on devices it has never seen before. Anthropic emphasized that there has not yet been a program that is completely out of control and truly autonomously replicating. However, the capabilities of intelligent agents in all aspects are still rising, and there is a high probability that there will be many variants that are beyond human control in the future. A sign worth paying attention to is: when the number of slow copying programs begins to increase, it often indicates that faster and more ruthless versions are coming. For the industry, this is both a clarion call to harden the system and a sober reminder - while trying to build a strong line of defense and block the spread of such programs, we must also do a good job in psychological preparation: they are likely not to disappear, but will live in our computing devices as a long-term thing. via AI News (author: AI Base)