(FORTUNE Magazine) – Hey! Guess what those wild and crazy Microsoft geeks in Redmond,Wash., are working on now? Statistical physics! Pretty cool, huh?

In the words of Jennifer Tour Chayes, she and her husband, Christian Borgs, both statistical physicists who recently joined Microsoft's rapidly expanding research group, are looking to "derive the observed behavior of gases, liquids, solids, and other states of matter from the underlying microscopic world of molecules, atoms, and electrons." One of their favorite topics is "independent percolation," a model that describes problems ranging from "the distribution of oil in a porous medium to the distribution of matter in the galaxy." Of course, "certain dependent versions of the percolation model, namely the integer random cluster models, are equivalent to the basic models of equilibrium statistical mechanics; namely, the Ising and Potts magnets." Got that?

Like you, perhaps, Chayes couldn't imagine what use Microsoft might have for her expertise when she was first contacted last May by her old Princeton University classmate, Microsoft's chief technology officer, Nathan Myhrvold. It's a lot easier to understand why Microsoft would invest, say, $1 billion in the Comcast TV-cable company or spend $425 million to buy WebTV Networks. For the company to drop something like a six-figure salary and stock options on Chayes seems a bit far-fetched. She's been on the payroll six months and is still not sure when--or whether--she'll deliver anything tangible to the company. When she recently met CEO Bill Gates, she commended him for investing in research that "won't pay off for 100 years."

At a time when corporations across the country are pulling back from basic research, Gates and Myhrvold are intent on building one of the all-time great research organizations--an R&D dynasty that people will mention in the same breath with such legendary idea empires as Bell Laboratories, IBM's Thomas J. Watson Laboratory, and Xerox's Palo Alto Research Center (PARC). In the past six years, Microsoft has quietly assembled 245 of the brightest researchers from around the globe, drawing from top universities and corporate rivals. Gates plans to hire another 400 researchers in the next three years, and build research centers around the world to complement the one in Redmond and another he and Myhrvold are setting up in Cambridge, England. The goal is as ambitious as you might expect. Says Gates: "The future of computing is the computer that talks, listens, sees, and learns. That is what is being created at Microsoft Research."

To convey his boss to that future, Myhrvold has populated Bill Labs with a lineup of Hall of Famers from the past. Now working there are the inventor of the laser printer (Gary Starkweather); the father of the VAX, the minicomputer that made Digital Equipment the world's No. 2 computer company for a time (Gordon Bell); the co-inventor of Ethernet and chief designer of the Xerox PARC computer that inspired the Macintosh (Chuck Thacker); the graphics genius who co-founded Pixar, the company that made Toy Story (Alvy Ray Smith); and the designer of the PARC technology behind Microsoft Word (Charles Simonyi). The Seattle Mariners have Ken Griffey Jr., Jay Buhner, and Edgar Martinez at the heart of their lineup. Not bad. But Microsoft Research is more like Murderer's Row, the Babe Ruth-Lou Gehrig-Tony Lazerri New York Yankees, of 1927.

Supporting this roster, plus 200 or so bench players like Chayes, probably costs Microsoft much less than $100 million a year. That's a whole lot less than Gates might spend on another alliance with a cable company. This side bet on research, however, may prove to be the gambit that distances Microsoft forever from the competition. Of the next four biggest independent American software companies--Computer Associates, Oracle, Novell, and Sybase--only Novell has a basic research group, and it possesses a skeleton staff.

The company that touts itself as Microsoft's archrival, Sun Microsystems, does have a lab. Its 100 researchers work on projects related to software and hardware, and the contrast with Microsoft is illuminating. Sun Labs restricts its researchers to tangibles like software for electronic commerce and software testing tools. Mathematical physics theorists need not apply. Says Sun Labs fellow Robert Sproull, an eminent computer architect who worked for many years at Xerox PARC: "We are not taking the big risks." CEO Scott McNealy, who contends that research centers "tend to be black holes for money," has kept the head count at Sun Labs constant over the past seven years, even as Sun's annual sales have climbed from $3 billion to $10 billion.

Gates is betting that approach is shortsighted. He and Myhrvold believe Microsoft stands a far better chance of taking the lead in technologies that will shape computing's future. And, says Myhrvold, focusing on the future can spin off products right now. (See "What's the Return on Research?")

If Microsoft Research does put Gates' imprint on the future, it will be an unlikely triumph. When Gates and Myhrvold created the unit in 1991, Microsoft was anything but the purveyor of soup-to-nuts software offerings we see today. Microsoft was the software industry's plumber. It derived most of its $1.8 billion in 1991 revenues from lowly MS-DOS; Windows had not taken off. Early in 1992, industry newsletter SoftLetter predicted Gates & Co. had little future in applications software. Instead, wrote the editors, "Lotus will dominate Windows spreadsheets, WordPerfect will still be top dog in word processing, and Borland will prevail in databases." That such a company would aspire to own a research organization akin to IBM's or AT&T's--i.e., a lab that produced mind-blowing research and great products--seemed laughably audacious. Yet Gates had always wanted a research division. "Even in the late 1970s," he says, "we were watching the work being done at Xerox PARC and talking about when we would be able to have a research group like that."

Assigning Myhrvold was perfect, if obvious, casting. Myhrvold's mother says Nathan told her at age 2 that he wanted to become a scientist. He graduated from high school at age 14, from UCLA at 19 with a bachelor's degree and a master's in geophysics, and from Princeton at 23 with a Ph.D. in mathematical and theoretical physics and a master's in mathematical economics. He once described his primary research interest as that portion of time when the universe was "about 10 -33 [exponent] of a centimeter, up to when it was about the size of a grapefruit. After that, it's all sort of history as far as I'm concerned."

He left a fellowship under Stephen Hawking at Cambridge University to try his hand as a software entrepreneur. Microsoft bought the business in 1986; by 1991, Myhrvold was running Microsoft's advanced development unit. He says its job was "doing things others agree are theoretically possible but have not been built before." Given a shot at organizing basic research--"attempting to do something only the investigator thinks is possible"--he jumped.

Myhrvold prepared a long memo for the board of directors, asking for $10 million a year to get started. In the remarkably prescient document, he argued that the personal computer industry had kept itself busy for years finding ways for little computers to mimic the functions and take over the jobs of big ones. But few such problems remained. Myhrvold wanted Microsoft to be at the forefront of the next generation of software by, for instance, making computers much more user-friendly--a quality mainframes never needed. "The only way to get access to the strategic technology is to do it yourself," he wrote.

Acknowledging that IBM, Xerox, and especially Apple had little new software to show for their investments in basic research, he outlined a structure to help Microsoft avoid the same fate. For starters, Microsoft Research would focus exclusively on software, and only on those areas that would so obviously benefit the business that Microsoft could not afford to leave them to academic researchers or other companies to pioneer. The first hire would be a research director, someone with enough prestige to wow prospective researchers, and with the management skills to run an organization that would total 60 people by the end of the third year.

Myhrvold and the director would then hire the best and the brightest, encouraging them to continue the research that had attracted Microsoft. Ideally, there would be groups of five to ten researchers, each group tackling some aspect of key issues Microsoft wanted to address. Researchers would not be micromanaged; rather, they would be encouraged to interact with one another. "It's a little like conducting a dinner party," Myhrvold explains today. "[You don't] interrupt the conversations and tell people what they should be saying and thinking. If you pick the right people to convene, more and better things happen than you could have planned."

The board gave the okay. Myhrvold anticipated that bringing in the first hires might be tough--in the memo, he worried that Microsoft risked being perceived as a "bunch of hackers writing for toy PCs." But he persuaded Carnegie Mellon's Rick Rashid, the mastermind of a state-of-the-art operating system called Mach, to sign on as director. Rashid, in turn, went after Dan Ling, an IBM research manager with several patents to his credit, as his deputy. Even though Rashid and Ling had been undergraduate roommates at Stanford, it took Rashid six months to recruit Ling.

After that slow start, Microsoft has succeeded in winning the best talent. To university researchers bored with the insularity of the ivory tower and the perpetual quest for grants, and to corporate researchers fleeing downsizing and growing pressures to develop products, what was once the land of DOS now more resembles Oz. Microsoft Research is full of stars from academe and refugees from bedraggled competitors like Apple, Silicon Graphics, and Digital Equipment, all drawn by Myhrvold's offer of research freedom and Gates' stock options. Microsoft has absorbed whole teams, like a trio of natural-language experts from IBM. On campuses, the company has proved so voracious a recruiter that after Princeton University graphics wizard Michael Cohen signed on, a Stanford colleague E-mailed him a rueful note: "Last one in academe, turn out the lights." Jim Morris, chair of Carnegie Mellon's computer science department, laments that Microsoft has lured five of its faculty in the past five years. "It gives colleges an Excedrin headache," he says.

Myhrvold now has 245 propeller-heads--a concentration of, say, 40,000 IQ points--onboard. The researchers are attacking a spectrum of problems that range from the mundane (creating better tools for software developers) to the futuristic if obvious (a post-Windows NT operating system code-named Millennium, that is self-configuring, self-adjusting, and fault-tolerant and lives in the network) to the simply weird (helping computers read the emotion behind your facial expressions so they can deduce what you'll need next). Almost all of the researchers work in Building No. 9 on the Redmond campus--just around the corner from Gates' own office.

Six years into the project, Myhrvold seems to have achieved his goal of hosting an IQs-on-steroids dinner party. Even the most arcane projects are spawning useful ideas. Take statistical physics. Gates balked last year when Myhrvold urged him to hire Chayes and Borgs--"This is the only proposal which made me think twice," Gates says. But researchers in another group--cryptography--quickly latched on to the couple (whose group is called simply theory) after they arrived on campus. The cryptographers thought that Chayes' and Borgs' statistical insights into phase transitions--they have studied the shifts that occur when, say, a liquid turns into a gas--might help them devise cryptographic sequences that at certain points change, producing coded messages and passwords that truly are unbreakable. Now, as cryptography and theory pursue their separate studies, they consult to see where their work intersects.

If you woke me out of a sound sleep and asked, 'What are you?' " says Myhrvold, "it is unlikely that I would blurt out 'manager.' " Still, it's his 1991 organization plan that allows for the dinner-party atmosphere. And it's his careful planning that so far has allowed Microsoft Research to work well with the company's product managers, who are always looking for the next product to push out the door. Says Cohen, who manages the graphics group: "It's part of your job. When you have something that's good, you look for people within the company to productize."

Two years ago, researcher David Kurlander developed a program that transforms the typed comments of people "chatting" with one another across the Internet into an interactive, real-time comic strip. Called Comic Chat, it creates a comic-book-style avatar for each person chatting. The typed words appear in bubbles over the character's head. Best of all, the character's facial expressions change according to the tenor of the chatter's words--type in "Bill, you naughty megalomaniac!" for example, and your character might scowl and appear to scream. When the Internet product group got wind of Kurlander's research, they grabbed it--and him. For the past 18 months he's been working in the product group, enjoying what he calls "a sabbatical" from research. And his work? It's now called Microsoft Chat 2.0 and is available from www.microsoft.com.

"We didn't want a situation like Xerox [PARC], where the research was decoupled from product design," says Gates. "[We want] people who are supersmart but also have a desire to see their work in use." So when Rashid and Myhrvold look for new hires, they seek out eggheads who don't want their life's work confined to prestigious scholarly journals.

That kind of practicality appealed to Eric Horvitz, Jack Breese, and David Heckerman. In 1992 they were co-founders of a startup called Knowledge Industries in Palo Alto. They were tinkering with artificial intelligence programs that might help diagnose diseases--or disorders in everything from copiers to the space shuttle. Then Myhrvold came calling. Says Horvitz: "We were offered the opportunity to bring our stuff to millions of people." The trio licensed their technology to Microsoft, moved to Redmond, and started the company's decision theory and adaptive systems group. Ignore the daunting name; Horvitz and company have created some smart applications that you may have already used.

Most artificial intelligence researchers have relied on "rules based" models. To oversimplify, they've tried to teach computers to solve problems in a step-by-step, linear way. "Printer won't print? Then try solution A. Still doesn't work? Then you must want B. Still doesn't work?..." Such an approach sometimes succeeds--often, however, it's the digital equivalent of getting a nitwit on the line when you call customer service.

Horvitz and his cohorts trashed "rules based" models, opting instead for a modular approach in which the computer asks the user a series of questions, cross-checks the answers with its data on previous solutions, narrows down the universe of possibilities, and suggests what it thinks is the most probable and least expensive solution--in less than a second. To get a sense of how, and whether, this works, check out the Troubleshooting Wizard on Microsoft's technical-support Web page at support.microsoft.com/support. (Hint: Click on "More Search Options," and then on "Natural Language search" before you try anything too demanding.) You'll tap into the same software Microsoft's technical-support staffers consult when you call in. According to a very satisfied Breese, more than 600,000 Web visitors used the software in August--hundreds of thousands more than might have ever used anything created by his old company. As Myhrvold is quick to point out, that's 600,000 customers Microsoft has served who otherwise would have required more expensive human intervention.

The Troubleshooting Wizard is a primitive example of how Research is trying to create "the computer that talks, listens, sees, and learns." More far-reaching are the efforts of the natural language and speech recognition groups to get computers to understand the way we write and speak. Steve Richardson, George Heidorn, and Karen Jensen are IBM veterans who fled Armonk for Redmond six years ago when Big Blue cut their funding. The three have spent their professional lives in a quest to make machines understand language. In Microsoftian terms, they're way old farts: Richardson is 44 years old, while Heidorn and Jensen are both 59.

Nevertheless, the trio's 22-person group is doing some of the coolest stuff in Redmond. One of its creations is MindNet, a linguistic database constructed with the contents of two dictionaries. With a built-in parser that breaks down the grammar of every sentence you write faster than your seventh-grade English teacher, MindNet automatically finds links between words that are similar in meaning. Using MindNet, a computer "knows" that in the sentence "We were watching the comet with a large telescope," "telescope" is linked to "watching," not to "comet." So far, the program recognizes seven million links between different kinds of words; now the researchers are going to pour in the contents of a couple of encyclopedias, giving the program a new set of smarts.

Gates explains that one possible use for all this is more accurate searching on the World Wide Web. Myhrvold points to the natural language group as an example of the way that near-term benefits shake loose from long-term research. The grammar-analyzing function that the team cobbled together worked so well that the Microsoft Word product group integrated it in Word 97, replacing a grammar checker the company had licensed. According to Myhrvold, that saved millions of dollars in royalties.

The charter of Microsoft's speech technology group, which has been run since 1993 by Xuedong Huang, a Carnegie Mellon veteran, is to help PCs understand the spoken word. Microsoft doesn't yet offer its own speech recognition program: Huang says that even though the group has made great strides in the past year, a computer running Microsoft's best prototype gets only 93% of the words right when addressed by a businessperson with no special training. Huang's goal is to combine his best technology with MindNet, thus creating a thinking machine that understands, and that can act on, a user's every spoken word.

Sounds great. But IBM and two smaller companies called Dragon Systems and Kurzweil have already wowed reviewers with simple speech recognition systems that allow users to dictate documents and direct applications by voice command. These inexpensive programs (all three offer versions for under $100) are expected to be popular items at Christmas.

Believe it or not, Huang is likely to keep his head. First of all, Gates seems to accept that investing in research is a long-term play. Second, Gates has got the just-in-case option: Just in case you can't invent it fast enough, buy it. After all, if you're Bill Gates, you can afford it. In August, Microsoft paid $45 million for a minority position in speech-recognition software company Lernout & Hauspie. Besides getting access to Lernout's technology, Gates will get a slice of this Christmas' sales after all--Lernout bought Kurzweil earlier this year.

Indeed, it is the background presence of Microsoft's cash that gives Microsoft Research the aura of something invincible. The company pays Myhrvold's senior researchers salaries in the low six figures, about what they might have earned in academia. It also, of course, gives them plenty of stock options. The researchers swear--they just swear--that options have nothing to do with their decision to come to Redmond. "What draws us here is intellectual curiosity," says Steven Shafer, another researcher who left Carnegie Mellon. "We take a risk--the stock market could crash." Adds Myhrvold, with a perfectly straight face: "Academics aren't familiar with stock options."

Still, it can't hurt to know what's possible should you abandon Ann Arbor for the siren call of Redmond. "I can honestly say stock options had zero influence on my coming here," says Cohen, the ex-Princetonian, "[but] it's great to win the lottery." Rick Rashid joined Microsoft in 1991; four years later he made $2 million on options. In May, Myhrvold himself sold 900,000 shares obtained through options, for a tidy profit of $104 million.

Numbers like those help explain why few academics expect Microsoft to have trouble expanding beyond 600 researchers over the next three years. Myhrvold is bending over backward to make sure Microsoft Research stays attractive in other ways dear to researchers' hearts. While it interferes with his dinner-party idea, he and Gates realize that at least some geniuses don't want to move to rainy Redmond. For these, he is willing to set up satellite offices. That's what won over Gordon Bell, the dean of the minicomputer, and Jim Gray, a legendary designer of databases. With five others, they work out of a smallish office in San Francisco. Bell offers advice on research projects--he's particularly enamored of a Microsoft effort to make videocameras a standard PC component--and doubles as Gates' Silicon Valley spy. He was the one who suggested that Microsoft purchase WebTV Networks, which is still based in Palo Alto.

With the opening of its Cambridge University facility last summer, Microsoft launched a global recruiting effort. It will spend $80 million over the next five years building and staffing the lab, and has promised to pump $16 million in venture capital into tech companies in the area. Says Myhrvold: "The reason to locate a lab someplace else is to draw on a source of people that you would otherwise not be able to hire. Research is about getting great people--no other aspect matters as much as that." For years, he and Gates had pursued Chuck Thacker, who, while at Xerox PARC, designed the Alto computer. A commercial flop, the Alto was the first personal computer that offered users an intuitive graphical interface--in many ways, it was the model for the Macintosh. Thacker didn't want to move from Palo Alto to Seattle--but once Microsoft decided to open up shop in England, he quickly signed on so that he could work with Roger Needham, a renowned Cambridge University scientist who is heading up Microsoft's British lab. Myhrvold says he and Gates are looking at sites in Europe and Asia as possible future campuses for Microsoft Research.

Their ambition couldn't be more at odds with the direction of most U.S. companies. In a recent paper entitled "The End of Unfettered Research," Andrew Odlyzko, a researcher at AT&T Labs (the piece of Bell Labs that didn't go to Lucent), dispassionately describes the state of basic research in the U.S. Unfettered research, he writes, is "almost totally gone from industrial and government laboratories, and is under pressure even in academia." According to a recent survey by the National Science Foundation, U.S. corporate spending on basic research, running around $6 billion a year during the 1990s, has declined as a percentage of revenues.

Reining in corporate spending on basic research isn't necessarily a bad thing. For decades IBM was the exemplar of doing well by doing good work in the lab; its researchers around the world, including those at the Thomas J. Watson Laboratory near New York City, invented the scanning tunneling microscope, Josephson junction circuits, high-temperature superconductors, and a host of computing innovations like the hard disk drive and SQL, the leading database language.

When CEO Lou Gerstner arrived in 1993, he cut R&D spending by $1 billion. Basic research projects that had no clear relation to products were decimated. Nevertheless, Big Blue research has continued to excel. This year IBM announced a method for using copper as a conductor on computer chips, while the researchers' prize supercomputer, Deep Blue, conquered chess champion Gary Kasparov.

So far, Microsoft Research has invented nothing comparable--a grammar checker and a better help desk don't exactly match up with a revolution in chip technology. Still, Myhrvold argues that Microsoft has an edge over IBM, since Big Blue must spread its research over both hardware and software. "If you look at their historical ability to effect products, I think you would find that IBM's record in using things from research in software is nowhere near as good as in hardware," he says. "I don't mean to insult IBM's software people. It's not their fault. They are further removed from the software market than we are."

For the moment, Microsoft seems to be beating IBM on the recruiting field. Jean-Claude Latombe, chairman of computer sciences at Stanford, says that when his graduates consider potential corporate employers, "everything I hear is that Microsoft is the most attractive."

In part, that's because while others are cutting back, Microsoft is spending more and more on Bill Labs. Says Gates: "The board has always supported my R&D plans. They recently had a presentation from Research, and their main feedback was simply that we should do more." When Myhrvold is asked whether there is any scenario under which he can envision Gates' pulling the plug on basic research, he responds, "Alzheimer's." He adds: "You're asking an odd question. Bill isn't going to pull the plug on research any more than he will pull the plug on Microsoft." And what of the fact that by pushing so hard on research Microsoft is moving in the opposite direction from much of corporate America? "It means we're either really smart or really stupid," Myhrvold says. "Whenever you're greatly at odds with the rest of the world, one of those two things is true."

Nobody's ever accused Microsoft of being really stupid.

REPORTER ASSOCIATE Alicia Hills Moore