Meet The AI Agent With Multiple Personalities

In the coming years, work It is expected to work more than people, including computers and smartphones. So far, They make a lot of mistakes to use a lot.

A new agent named S2, a new agent, created by Simular AI, combines the frontage models with specialized models for use of computers. The agent can use the most modern performance over tasks such as using applications and manipulating and manipulating, and suggestions that have become different models in different situations can help the agents move forward.

“Computer use agents are different from large language models and coding,” said Codounter and COKUL CEO. “This is a different kind of problem.”

A powerful general purpose AI model like Openai’s GPT-4O or anthropic’s Claude 3.7 is used to use the best to complete the best to perform tasks such as translating websites.

Lee, which is a researcher in Google Deepmind, explains that in 2023, in the planning of large language models, but not good to recognize the elements of the graphical user interface.

The S2 is designed to learn from experience with an external memory module that uses these articles to record actions and user feedback and improve future actions.

Especially in complex positions, S2 is better performing better than another model OsWorlda criterion that measures an agent’s ability to use a computer operating system.

For example, S2 can complete 34.5 percent for beating, beating, Openai operatorAble to complete 32 percent. Similarly, S2, a benchmark for agents using smartphones, the next best agent 46 percent in Smenfmark.

Victor Zhong, one of the creators of Waterloo University and OSWORLD, believes that the belief that Big AI models can combine the visual world and feel the training information that helps feel the graphical user interface.

“It says that the agents will help the guis with higher accuracy,” he says. “Meanwhile, before such fundamental progress, the most modern systems will be similar to many models for combining many models to patch the restrictions of single models.”

To prepare this column, order to order flights and use it a bit to mix Amazon for transactions, and it looked better than some open sources I worked, including Autogenic and vimgpt.

But even the most intelligent AI agents, it seems, it seems, still worried about the outside and sometimes exhibits the only behavior. In one case, when asked S2 when asked to find contact information for researchers behind Osworld, the agent was closed to a loop between the project page and the conflict of OSWORLD.

OSWorld’s prices show why agents stay more than the reality so far. Although people can complete 72 percent of Osworld positions, agents make up 38% of complex tasks. He said Benchmark could complete only 12 percent of the best agent positions when submitted in April 2024.