The thing that makes computers a huge pain for everybody, says Pedro Domingos, an associate professor of computer science at the University of Washington, is that you have to explain to them every little detail of what they need to do. “It’s really annoying,” Domingos jokes. “They’re stupid.”

That’s why Domingos is taking part in CALO, a massive, four-year-old artificial-intelligence project to help computers understand the intentions of their human users. Funded by the Defense Advanced Research Projects Agency (DARPA), and coordinated by SRI International, based in Menlo Park, CA, the project brings together researchers from 25 universities and corporations, in many areas of artificial intelligence, including machine learning, natural-language processing, and Semantic Web technologies. Each group works on pieces of CALO, which stands for “cognitive assistant that learns and organizes.”

Adam Cheyer, program director of the artificial-intelligence center at SRI, explains that CALO tries to assist users in three ways: by helping them manage information about key people and projects, by understanding and organizing information from meetings, and by learning and automating routine tasks. For example, CALO can learn about the people and projects that are important to a user’s work life by paying attention to e-mail patterns. It can then categorize and prioritize information for the user, based on the source of the information and the projects to which it is connected. The system can also apply this type of understanding to meetings, using its speech-recognition system to make a transcription of what’s said there, and its understanding of the user’s projects and contacts to process the transcription intelligently into to-do lists and appointments. Finally, a user can teach CALO routine tasks such as purchasing books online and searching for bed-and-breakfasts that meet specific criteria. CALO can interact with other people, taking on tasks such as scheduling meetings, coordinating among people’s schedules, and making decisions, such as deciding to reschedule a meeting if a key member becomes unable to attend.

“It’s an amazingly large thing, and it’s insanely ambitious,” Domingos says. “But if CALO succeeds, it’ll be quite a revolution. Even if it doesn’t, so much good research is happening under it that it will still have been worthwhile.”

The goal is to build an artificial intelligence that can serve as a personal assistant for someone–not something with a rigid structure within which it can be helpful, like the animated paper-clip assistant featured in Microsoft Office products, but a system that can learn about a user’s environment and needs, and adapt to them, without having to be programmed anew by engineers. “What’s different and has never been done before in this way is the truly integrated approach of bringing all of these technologies and all of these capabilities into a single system,” says Cheyer. “It takes a system of this size to give you something that can understand and organize so much information.”

The project might seem broad in its goals, but the researchers believe that ultimately, the system will benefit from multiple technologies working together. Consider the meeting-transcription function, says William Mark, vice president of the information and computer-science division at SRI. Even the best speech-recognition systems would have trouble producing an accurate transcript of a meeting unassisted, he says, but “in our context, because of information management, CALO has deep and rich knowledge about who are the people in the room, and what are the documents and phrases and slang used in context.”

Since CALO has many learning systems, one challenge is integrating them so that CALO has a consistent structure for information that it can use to make decisions based on the noisy, uncertain data that it extracts from its various interactions. Domingos and others have been working on a probability consistency engine, which unifies two traditional approaches to artificial intelligence: logic and probability.

Alan Qi, an assistant professor of computer science at Purdue University, who is not involved with CALO, says that the unification of logic and probability is an important endeavor for the field of artificial intelligence. Combining these two approaches, Qi says, is far better than using either alone. Probabilistic approaches can handle noise and uncertainty well, while a logical structure is best for handling meaning.

Although CALO’s approach is very far-reaching, SRI has made a version, called CALO Express, that boils down some of the features of CALO that are almost ready for deployment. CALO Express is a lightweight version of the real deal that integrates with Microsoft products such as Outlook and PowerPoint. Cheyer says that it includes parts of the three main features of information management, meeting assistance, and task management. He says that CALO Express is now being evaluated for use at DARPA. While it’s uncertain whether CALO Express will become a commercial product available outside of the military, there is still hope that the average person may get access to technologies of this type. The research has already produced a few products, such as Smart Desktop, which is an information-management system that spun off of the task-tracer project done by Oregon State University as part of CALO. Radar Networks, makers of the Semantic Web product Twine, has also worked on some of CALO’s semantic underpinnings. (See “The Semantic Web Goes Mainstream.”)