(FORTUNE Magazine) – ALMOST EVERYBODY has heard by now about artificial intelligence: computers that will be able to do things that dullard humans can already do instantly, such as spot a face in a crowd, understand speech, or look out the window and guess whether it will rain -- tasks that depend on recognizing patterns. Various AI machines have been soaking up millions from big spenders like Japan's MITI and the U.S. Defense Department. But did you know that you can take another route to AI with a piece of software as cheap as $99 that will run on your PC? This vastly different approach is called neural net technology, since it vaguely mimics the brain's complex network of neurons, the cells that transmit and store the nervous system's messages. Neural nets do have their hitches. Like people, they arrive at acceptably accurate decisions in mysterious ways. They can't tell you how they solve a problem, and it's unclear how big a task they can take on. So far no one really understands either their limitations or their ultimate usefulness. Another difficulty: Neurocomputers have barely reached the fledgling stage, so most applications must be simulated on ordinary digital computers. That's what that $99 software lets you do. Fortunately, researchers with impressive credentials are being drawn to the field. In 1986 Federico Faggin and Carver Mead, whose work revolutionized the design of semiconductor chips a decade ago, started a company called Synaptics in San Jose, California, to devise a completely new type of chip for neural computers. Even they caution that neural nets are in their infancy. ''This will be an art for a long time before it becomes a science,'' says Faggin. If it succeeds, neural net research could lead to a generation of machines thousands of times faster than today's computers, capable of interpreting speech, vision, and data in ways that are impossible now. Neurocomputers even seem to be able to spot patterns that people can't. Because they can identify those patterns without special programming, they may turn out to be an improvement over so-called expert systems, a rival branch of artificial intelligence that requires elaborate rules yet can rarely deal with unfamiliar information. If a neural net is faced with the problem of figuring out whether dog ownership has any effect on a person's creditworthiness, it eventually teaches itself the answer (see box). An expert system can't do that; the programmer has to tell it whether owning a dog is significant. Automakers, oil drillers, chemical producers, and financial service companies have had mixed but generally promising results with neurocomputing. In the mid-1980s, for instance, occasional noisy blower motors turned up among those that Siemens, the giant West German electrical equipment maker, was producing for Ford car heaters in Europe. To identify the defective ones, the company did the obvious: Workers turned on the motors and listened, but that was so tedious that performance dropped off quickly. Siemens tried to program a traditional computer to recognize the offending mix of sounds, but couldn't get it right. A year and a half ago the company tested a neural net device developed at the Siemens research lab in Princeton, New Jersey. ''It turned out to be correct more than 90% of the time,'' says project leader Wolfgang Feix. Siemens now checks all blower motors with neural nets. IN THE PAST THREE YEARS hundreds of tiny neural net companies have sprung up. Few are profitable, and so far the lot of them resemble a fantasy more than an industry. Even two well-regarded companies, both publicly traded, have had financial woes. Nestor Inc. of Providence was co-founded in 1975 by Leon Cooper, a Nobel Prize winner in physics. Nestor has never made a profit; last year it lost $2 million on revenues of $500,000. Excalibur Technologies Corp., in Albuquerque, New Mexico, went bankrupt in 1985, five years after it was formed, but has reorganized and expects to make about $100,000 on sales of $1 million this year. Revenues of the entire neural net industry in 1988 came to $20 million -- about $5 million in software and services and $15 million from traditional computer hardware that speeds up neural net simulations, says Tom Schwartz, a Mountain View, California, analyst. Schwartz and others predict that industry revenues will more than double in 1989 to about $50 million. The business has stayed small in part because some applications simply haven't worked. Forget using these things to beat the stock market, for example. The Kidder Peabody brokerage firm tried to get a neural net to detect price and volume patterns in stock trading as a guide to future investment decisions. ''We didn't have much success,'' says Paul Sabota, head of a Kidder analytical trading group in New York. ''The data were extremely noisy.'' The biggest commercial success for neural nets has been in evaluating loan applications. ''That's about the only thing making any money,'' says Robert Hecht-Nielsen, a mathematician who founded HNC of San Diego in 1986. Companies such as American Express that need to appraise credit risks say they are using neurocomputers but won't give details, apparently for competitive reasons. The new technology has one big fan in the oil field. Before Arco ''fractures'' a field, injecting fluid into well holes at high pressure to crack rock formations and improve the flow of oil, the company first develops huge computer models of it. Those models help engineers interpret data from sensors in the ground that tell them how the fracture is proceeding. Some models take hours to run, however, which makes them almost useless if a problem develops while fracturing is under way. Richard Stoisits, Arco's principal production engineer, used data stored in a mainframe computer to train a neural net on the characteristics of the Prudhoe Bay field. ''It can retrieve the answer almost instantly,'' says Stoisits. He says this ''spectacular technology'' now identifies some patterns better than he can: ''It's almost like magic.'' A chorus of small success stories can be heard from dozens of other companies and organizations. Ford Motor reports that neural nets can be trained to spot faulty paint finishes. Better than a human inspector? ''Better than a tired one,'' says Shaun Devlin, manager of vehicle electronics at Ford's research facility in Dearborn, Michigan. He's hopeful that neural nets will help diagnose engine problems as well. A major law enforcement agency is using a neural net to see whether the psychological profile of people who commit a particular type of crime is changing. GTE is installing a net to track variations in heat, pressure, and the chemicals used to make fluorescent bulbs; that will help determine the optimum manufacturing conditions. Martin Marietta has trained a robot arm to recognize and pick up pallets even when they are at cockeyed angles. Airports in Los Angeles, San Francisco, and London have contracted with SAIC, a San Diego hardware and software company, for neural nets that can help tell the nitrogen in a bomb from that, say, in cheese. Setting up a neural net is often a frustrating, trial-and-error process. For example, if the operator tries to get the net to learn too much too soon it won't respond, but if the training rate is too slow the process takes forever. ''The tools you get from a neural net company are not sent out ready to run,'' complains Tom Ansusinha, manager of advanced planning for information systems at Navistar. ''You call up the company and tell them you want to know how to adjust the learning rate. No one can tell you how. It becomes very time consuming.'' WHAT SHOULD a company do if it's interested in experimenting with neural nets? Devlin at Ford got his start with a personal computer and a $99 software package called Brainmaker from California Scientific Software in Sierra Madre. ''We didn't feed it any big problems, but it was fine as a learning tool,'' he says. A manager at that law enforcement agency recommends a $199 starter package called NeuroShell from Ward Systems Group in Frederick, Maryland. ''It's ideal for morons,'' he says. ''I can use data already in my personal computer, and it creates profiles that I would never see in looking through a thousand files myself.'' Bigger-ticket ways to get your feet wet in nets include a five-day course given by HNC that's designed to teach an expert in, say, credit evaluation how to devise elaborate systems for his company. HNC and SAIC offer software and $17,500 to $25,000 accelerator boards that greatly speed up neural net simulations on PCs. Some of the most impressive systems are proprietary, with secret wiring arrangements and learning formulas. Nestor devised one that could read as many kanji characters as a well-educated Japanese. The company's equally impressive list of customers includes Ford, Morgan Stanley, and BancTec, a Dallas-based service company for banks that plans to use the technology to read handwritten numbers on checks. For a minimum of $7,500, Nestor licenses a complete system for a particular application, with a money-back guarantee if it doesn't work. President Michael Buffa blames high research costs for years of losses, but says Nestor has begun to emphasize marketing and expects to become profitable % sometime this year. The stock, as low as $2 in 1987, has climbed to around $5. After Excalibur went bankrupt in 1985, a subsidiary of Nippon Light Metals invested in the company, paid for further research on the technology, and used it in its factories, says President Richard Duddy. Derek Stubbs, a Vicksburg, Michigan, consultant on neural nets, calls Excalibur's systems ''excellent.'' Excalibur handled the Prudhoe Bay simulation, which Stoisits of Arco says would have been virtually impossible using the publicly available formulas for training neural nets. Three years ago Excalibur stock was trading at 87 cents; now it is over $4. Before anyone makes big bucks in the neural net business, a few major hurdles have to be cleared: devising ways to present information in digestible form to neural computers and developing special chips and other hardware devices to speed up processing by a thousandfold or more. At AT&T Bell Labs in Holmdel, New Jersey, researchers Lawrence Jackel and Richard Howard produced a machine that can read handwritten zip codes with 99% accuracy. Writing computer code was fairly easy, says Jackel, but reducing the amount of visual information the computer had to process was far more difficult. ''You're trying to do what evolution took millions of years to accomplish,'' says Jackel. PERHAPS THE MOST ambitious work is under way at Synaptics, where Mead and Faggin are developing special chips that act as electronic eyes and ears. By using neuronlike arrays of tiny silicon sensors, their electronic retina can capture a moving image of a face, and by performing a billion operations a second reduce it quickly to a set of jagged lines that a computer can analyze. Faggin is pleased with progress so far, but still not satisfied. ''We are years away from a product we can release,'' he says. It's unlikely that a single, all-purpose neural net machine will emerge. Problems involving speech, for instance, seem to respond best when neurons are wired differently than for other tasks. Nor are neural nets likely to replace existing digital computers. The nets may serve as eyes and ears for them, says Caltech biophysicist John Hopfield, a leading researcher in the field. Neural nets may never steer cars around potholes on a freeway, as one neurocomputer company head extravagantly predicts. Even so, says Devlin of Ford, ''if all they do is solve a lot of little problems well, they're still worthwhile.''