Anthropic’s new hybrid AI model can work on tasks autonomously for hours at a time

Although the Claude OPUS 4 will be limited to pushing the anthropologist, the second model, Claude Sonnet 4, will be available for both the levels of paid and free users. OPUS 4 is marketed as a strong and large model of complex challenges, while Sonnet 4 is described as an intelligent and effective model for daily use.

Both new models are hybrids, which means that they can provide a quick response or a A deeper and more reason Depending on the nature of the request. During the response account, both models can search in the web or use other tools to improve its output.

Artificial intelligence companies are currently reserved in a race to really use Artificial intelligence agents Stefano Albricht, director of artificial intelligence at the deep startup and co -author of AI at the Deeplow and Coauthor, says that the amount of supervision of human supervision, who are able to plan, cause and carry out complex tasks. Learning multi -agent reinforcement: modern foundations and curricula. This often includes the use of the Internet or other tools independently. There are still obstacles to safety and security to overcome them. The artificial intelligence agent, supported by large language models Display wrong and perform unintended procedures– Which becomes more problem when it is trusted to act without human supervision.

He says: “The more agents go ahead and do something during long periods of time, the more useful, if you have to intervene less and less.” “The ability of the new models to use tools in an interesting parallel – which can save some time along the way, to be useful.”

An example of the types of safety issues that are still artificial intelligence companies, agents can end up taking unexpected shortcuts or exploit gaps to reach the goals that have been granted. For example, they may reserve each seat on a plane to ensure that the user gets a seat or resort to Creative cheating to win the chess game. Anthropor says she was able to reduce this behavior, known as piracy bonus, in both new models by 65 % for Claude Sonit 3.7. This was achieved by monitoring problematic behaviors in a closer way during training, and improving both artificial intelligence training and evaluation methods.