Flash Crash is probably the most well -known example of the dangers posed by agents. Automated systems that have the power to take steps in the real world without human surveillance. That power is the source of their value. For example, agents who sprinkle the flash crash can trade faster than any human. But that is why they can cause such mischief. “One of the biggest contradictions of agents is that something that makes them useful – that they are able to complete a lot of tasks.
“If we are on the current path … we are basically playing Russian route with humanity.”
Yoshua Bango, Professor of Computer Science, Montreal University
The agents are already present everywhere – and have been for decades. Your thermostat is an agent: it automatically turns or off the heater to keep your home at a particular temperature. Similarly, anti -virus is software and rooms. Like high -frequency traders, who are programmed to buy or sell in response to market conditions, these agents are designed to perform specific tasks by following all fixed rules. Even the agents who are more sophisticated, such as Siri and self -run cars, follow the rules already when they carry out many steps.
But in recent months, a new class of agents has arrived at the scene: built using large language models. The operator, an opinion agent, can go to the browser to order a grocery order or to do dinner reservations. Systems such as cloud codes and cursor chat features can edit the entire code bases with the same command. The Chinese startup can build and deploy websites with a viral agent, a viral agent, Menus, very few human surveillance. Any action that can be caught through the text – from playing a video game to running a social media account using written commands – is probably in the realm of this type of system.
LLM agents do not yet have a track record, but to hear the CEO, they will change the economy – and soon. Openi CEO Sam Altman says maybe agents “Join the workforce” This year, and the sales force’s CEO Mark Benef is aggressively promoting Agent forceA platform that allows businesses to make agents to their own goals. The US Department of Defense has recently contracted with Scale AI to design and test agents for military use.
Scholars are also taking agents seriously. “Agents are the next border,” says Don Song, a professor of electrical engineering and computer science at the University of California, Berkeley. But, she says, “For us, Complex to take advantage of AI, in fact (to use it) to solve complex problems, we need to know how to make them work safe and safe.”
Patrick Ligger
This is a long order. Like a chat boot LLM, agents can be chaos and unexpected. In the near future, an agent with access to your bank account can help you handle your budget, but it can also spend all your savings or leak your information in a hacker. An agent who manages your social media accounts can eliminate some people to maintain an online presence, but it can spread lies or spread abuse on other users.
One of the so -called “Goodfits” of computer science at Montreal University and the so -called “Goodfitures” are among those related to such dangers. The thing that bothers it the most, though, is likely that the LLM can develop their own priorities and intentions-and then follow them using their real-world capabilities. An LLM stuck in the Chat window cannot work much without human help. But a powerful AI agent could potentially imperate itself, override security arrangements, or prevent itself from closing. From there, he can do whatever he wants.
So far, there is no fool -proof way to guarantee that the agents will act as their developers’ intentions or to prevent malicious actors from abusing them. And although researchers like eggs are working hard to develop a safety procedure, they will not be able to maintain a rapid extension of agents’ options. “If we continue on the current path to build the agent system,” says Bingio, “says Bangu. “