Is AI becoming mainstream due to Microsoft? Matjaž Gams Department of Intelligent Systems Jožef Stefan Institute (and Ljubljana University) Jamova 39, 1000 Ljubljana, Slovenia Tel: +386 61 1773900; fax: +386 61 1251038 e-mail: matjaz.gams@ijs.si http://www2.ijs.si/~mezi/matjaz.html Abstract For decades AI promised to radically change the world of computing by introducing true intelligence to machines. The optimism more often resulted in disappointments than in true breakthroughs. But finally, intelligent systems are becoming mainstream through the Internet and Microsoft. This paper concentrates on the Microsoft AI research and applications. 1 INTRODUCTION Years ago, there were complains that computer systems lack even the most primitive and most often needed semi-intelligence. For example, if a user mistypes or uses a sequence of commands several times, every human knows what is the intended action of the user. At that time, none of the commercial computer systems had such capabilities. This seems strange because for decades AI has been building user-friendly systems that were capable even of communication in natural language (Davis 1998). Systems were capable of providing advice when asked in natural language about desired functions, e.g. "How to delete all backup files?". But these were research prototypes never applied in commercial applications. Twenty years later, similar improved systems revived as agents (Etzioni, Weld 1995; Gams, Hribovšek 1996) first implemented again as research prototypes. But until recently, even the most stupid mistakes had to be corrected by users without any help from commercial system. That is time-consuming and costly. Many studies have shown that users lose significant amounts of time on computer not doing the work they are supposed to. Older AI research programs aimed at achieving human-type intelligence by using pure logic. Now, several serious researchers think that true human-level intelligence can not be achieved on computers without radically new approaches (Sloman 1992, Penrose 1994, Edelman 1992). Currently, there are three AI divisions: The branch insisting on the formal old AI research is named "strong-AI". The new human-oriented or towards-cognitive-science oriented AI is called "new AI". The most productive in the sense of practical applications is the "engineering AI", or the "intelligent-systems" branch. The intelligent systems approach (Goonatilake, Treleaven 1995 ) is based on old and often well verified AI techniques, implemented in and adapted to real-life domains. A typical example would be the old alpha-beta search algorithm, implemented with several new additions including different types and levels of knowledge, and applied in the IBM Deep Blue supercomputer that defeated Garry Kasparov. The design team was awarded by the $100.000 Fredkin Prize and the Allan Newell Research Excellence Medal. Although Garry later insisted that he was actually the better player losing the psychological war (some analyses of the games confirm his claim), the result astonished experts and laymen alike. This was the latest public and media breakthrough. The practical breakthrough came with intelligent systems and agents on the Internet. In this paper, the emphasis is on the Microsoft AI (Hedberg 1998). 2 MICROSOFT'S RELATION TO AI Bill Gates (http://www.microsoft.com/billgates/) has been inspiring the growth of the Microsoft empire (http://www.microsoft.com) by forseeing forthcoming software systems. In several interviews already accompanying the Office products, he explained his new vision of the computer of the future - an intelligent computer assistant, a kind of secretary, capable of communicating and executing simple mundane tasks. The new system will be capable of talking, listening, seeing, and will have other anthropological features like faces capable of expressing gestures. The first simple implementation of this computer assistant is the Office Assistant, described later in the paper. Microsoft has long-lasting inspirations for "practical" AI. The Palo Alto research inspired Job's Apple and Microsoft Windows. In 1994, Steve Ballmer, until recently Microsoft's executive vice president, addressed the American AI conference by introducing several simple (or maybe even primitive by today's standards) intelligent agents such as Tip Wizard in Excel and Liz in the Magic School Bus. The ancient Microsoft agent program that we studied in our laboratory, Bob, was a children version with a barking dog. But all these systems had the same motive, the same vision: how to solve the problem addressed in the beginning of this paper. How to help users in positions when they approximately know what to do; however, are puzzled by extreme rigidity (stupidity) of classical computer system. The idea is to humanise computers (Maes 1997), to enable more productive work, and to eliminate the frustration produced by complicated technical details. 3 MICROSOFT RESEARCH Microsoft has a tendency of hiring young geniuses. These people bring new, fresh ideas into the Microsoft campus. The other group providing the bones and flesh of new products is based on already established designers that introduced major new computer systems. In Microsoft labs there are (were) Nathan Myhrvold, a theoretical physicist previously working with Stephen Hawkings finishing his PhD at 23; Gordon Bell, the constructor of Digital's Vax; Chuck Thacker, coinventor of Ethernet and major designer of the Xerox PARC Alto computer; Alvy Ray Smith, a cofounder of Pixar; three AI scientists from the Palo Alto group: Jack Breese, David Heckerman, and Eric Horvitz. Recently, Fayyad, one of the leading AI researchers in knowledge discovery and data mining, joined the research group just coming from the Caltech Jet Propulsion Lab. Microsoft's braintrust was recently based on four top people: Ballmer, Myhrvold, Higgins and Maritz. In July 1998, Bill Gates named Steve Ballmer president of Microsoft. Gates (CEO of Microsoft) will focus his efforts on new technologies and future products (http://www.microsoft.com/BillGates/billgates_l/s peeches.htm, Seattle, WA, June 2, 1998): "I'm really a big believer in research and the payback that it has. And I'm actually a little surprised how little the software industry has invested in research. The payoff is very, very clear. The need is very, very clear. In the information age, this ability to process information, to let people have access to the things they care about, it is the most exciting problem of all." Currently (July 1998), Rick Rashid is the head of the Advanced Technology and Research Group. His introduction on the Internet (http://www.microsoft.com/presspass/exec/rick/ric kspeech.htm): "Basically, everything Microsoft does these days, all of its products, have some influence from a research organization, whether it's actually the product itself, the features of the product or the technology used to create the product." While most corporations like IBM are decreasing their research and development departments, Microsoft is intensively increasing its up. In 1998, the research group has around 300 stuff with a $100-$200 million. Microsoft is seeing his perspective in developing cutting-edge products. To do this, Microsoft wants to build a brain trust, to increase research, to increase production of cutting-edge ideas. Before concentrating on pure SW AI research we present an overview of Microsoft's new ideas: new (versions of) operating systems, networks, word processors, graphics, Office etc. methods and are becoming more and more userfriendly. One of the basic ideas here is rather old to use the idle time of computer to improve its performance and to serve better. Besides operating systems and Office, other areas include multimedia, the Internet, networking, natural language speech and recognition, new tools etc. An example could be automatic personalised filtering on the Internet or intelligent information retrieval. Figure 1: Our Ema intelligent employment agent includes the freely available Microsoft agent dynamically speaking in English. internet tools and systems handheld PC's Internet telephones smart cars intelligent household machines like washers intelligent kitchens intelligent houses WEB TVs advertisements (movies, games etc.) banking, ecommerce. Several of these systems have already been implemented usually first in USA. Such an example is the WEB TV, introduced around half a year ago. Basically, one attaches something like a keyboard to a TV, something like the old Sinclair's microprocessors, and one can surf the Internet. Another example would be the hand-help or palmsize PC with Windows CE. 4 AI RESEARCH In Microsoft, there are many groups closely related to AI: Decision Theory and Adaptive Systems, Natural Language, Speech, Vision, and projects like the Millennium distributed operating system including self-tuning and self-configuration. Previous first versions like Windows Plus! already included self-tuning and self-checking, but these new versions include more and more advanced AI For example, Microsoft Explorer is competing hard (some say even unfair) with Netscape. To provide better services, new functions are added. One is the personalised WEB services, already existing on the end of the MSNBC site (http://msnbc.com/news). This site has 300.000 visits daily. Electronic commerce is also added (IntelligentCrossSell). Usama Fayyad is working on these features. The Internet Microsoft agent is freely available (http://www.microsoft.com/msdn/sdk/inetsdk/ help/msagent/agentdl.htm). It dynamically speaks English, has several visual forms and voices. This agent (se Figure 1) was also implemented in our intelligent agent employment system EMA (http://www-ai.ijs.si/~ema/) described elsewhere, e.g. in (Gams et al 1998). The recent generation of agents started with wizards, e.g. those helping in troubleshooting. In order to help users, the agent has to "understand" system functions, the problem and user's wishes. There are some 70 Web-based user-troubleshooter agents providing on-line help for different systems. With this, the users get on-line help 24 hours a day. Of course, the number of calls to the Microsoft stuff didn't decrease, but userfriendliness of Microsoft in troubleshooting problems substantially increased. The current plan is to introduce even more wizards into Windows 98 and NT 5. The troubleshooting wizards were designed by the Microsoft Belief Network tool. It consists of user interface, memory management and other functions enabling quick construction of a wizard for a specific problem domain. The system in a good demo version is available free for academic institutions at http://suppport.microsoft.com/ support/tshoot/default.asp. Bayesian statistics (networks) is basically one of the two main branches in statistics; the other is classical statistics. Bayesian statistics provides an explicit mechanism for combining prior knowledge with data. Intelligent agents and intelligent systems as well as older expert systems (Goonatilake, Treleaven 1995) are not defined by the computational method but with the functions they provide; therefore, Microsoft troubleshooting systems are intelligent systems. Microsoft is attracting excellent researchers not only by good working conditions but also by the satisfaction anybody gets when his/hers ideas get copied and sold in millions. Literature [1] R. Davis, What Are Intelligence? And Why? AI Magazine, Spring, pp. 91-110, 1998. [2] G. Edelman, Bright Air, Brilliant Fire, On the Matter of the Mind, Penguin Books, 1992. [3] O. Etzioni and D. S. Weld, Intelligent Agents on the Internet: Fact, Fiction, and forecast, IEEE EXPERT, Intelligent Systems & Their Applications 10, No. 4, pp. 44-49, 1995. [4] M. Gams and B. Hribovšek, An IntelligentAgent Interface for Operating Systems, Proceedings of ITHURS'96, Spain, pp. 167-174, 1996. Figure 2: Office Assistant is with 70.000 sold pieces daily the most often sold Microsoft agent and also agent of any kind. Another well-known product of this family is Lumière, the electronic butler. The modified version of Lumière is Microsoft Assistant, a friendly assistant in Microsoft Office (Figure 2). Assistant Office replies to user questions or orders in plain English. The reply typically consists in a form of several possible options based on Assistant's knowledge about similar problems and the context of the user. Afterwards, the user chooses one of the possible replies, and Assistant shows the sequence of actions to perform the task. Assistant exists in several graphical forms from Shakespeare to the Earth. Each icon expresses gestures to indicate the state of the system. 5 DISCUSSION Microsoft agents are among the most often used agents worldwide. They deal with mundane tasks, and try to make systems more intuitive and easier to use. These systems are very helpful because they implement user-friendly AI methods for the most needed improvement in human-computer communication and understanding. [5] M. Gams, A. Karalič, M. Drobnič, V. Križman, EMA - An Intelligent Employment Agent, Proc. of the Forth World Congress on Expert Systems, Mexico, pp. 57-64, 1998. [6] S. Goonatilake and P. Treleaven (ed.), Intelligent Systems for Finance and Business, Wiley, 1995. [7] S. R. Hedberg, Is AI going mainstream at least? A look inside Microsoft research, IEEE Intelligent Systems, March/April, pp. 21-25, 1998. [8] P. Maes, Humanising the global computer, IEEE Internet Computing, July/August, pp. 10-30, 1997. [9] R. Penrose, Shadows of the Mind, A Search for the Missing Science of Consciousness, Oxford University Press, 1994. [10] A. Sloman, The Emperor's Real Mind: Review of the Roger Penrose's The Emperor's New Mind: Concerning Computers, Minds and the Laws of Physics, Artificial Intelligence 56, pp. 335-396, 1992. S. Hamilton: Inside Microsoft Research, Computer, pp. 51-58, January 1998.