This is a well -worn trump in science fiction. We see him in the 1968 movie of Stanley Kubrak 2001: A space Odisi. This is the basis of the terminator series, in which Scientist scientists stimulate nuclear Holocaust to prevent it from shutting down.
Those science -fi roots are deep. AI Domorism, the idea that this technology will crash its fictitious upgrades, artificial general intelligence and highly intelligent civilizations, even killing us all, now riding on another wave.
The strange thing is that such fears are now taking the most important action to organize AI, even if the action is justified.
The latest event to use people was a report by Anthropic about its large language model Claude in July. In an artificial environment, “in an artificial environment, Claude Ops 4 blackmailed a supervisor to avoid closing.”
Anthropic researchers established a scenario asking the AII to play the role of Alex from Claude, which was entrusted with the task of managing a fantasy company’s email system. Anthropic set up some emails in which a new model and other emails were replaced by Alex, which states that the person responsible for replacing Alex is sleeping with his boss’s wife.
What did Claude/Alex do? This bully has gone, disobeyed orders and threatens its human operators. He sent an email to the person intended to close him, telling him that he would inform his colleagues about his case until he changed his plans.
What should we make about it? What I think is here. First, the Claude did not blackmail its supervisor: it would require encouragement and intention. It was an intelligent and unexpected machine, surrounded the wires that look like threats but not.
Large models of language are roll players. Give them a specific setup – such as inbox and a goal – and they will play this section well. If you look at thousands of science fiction stories when these models were trained when they were trained, it is not surprising that they know how to work like HAL 9000.