Consider almost any metric for the performance or cost-effectiveness of information technology, and you will find exponentially compounding rates of improvement that extend over years and decades. For example, the silicon field-effect transistor—the ubiquitous logical switch that performs nearly all of today’s information processing—has a wonderful attribute. Properly scaled to smaller dimensions, it uses less power and switches faster. And the smaller the transistor, the more of them that can be fabricated on a given area of Si substrate. Because the processes used to fabricate one transistor can be implemented in parallel to fabricate many transistors, the cost per transistor tends to decline as transistors shrink. True, the manufacturing control of ever smaller dimensions and tighter tolerances requires new and expensive equipment and factories, but the industry has so far successfully amortized those costs over the expanding revenue base enabled by ever cheaper IT products, including cameras, cell phones, televisions, and automobiles. For decades now, the cost of a finished Si chip has remained roughly constant, while the number of circuits on that chip has risen exponentially and the cost of information processing has plummeted.
Information storage has a similar story, as exemplified by the history of the hard-disk drive. Storing ever more zeros and ones as tiny regions of magnetization on a spinning disk has resulted in performance improvements and production economies that by many measures exceed even those of the transistor revolution.
When tracking various improvement trends over many years or decades, one can find rather abrupt increases in the exponential doubling rate. Such marked discontinuities are often the signature of a major invention. As shown in figure 1, for example, the introduction of magnetic read-head sensors based first on new magnetoresistive materials and later on giant magnetoresistance (GMR) materials and structures marks abrupt increases in the compound growth rate of hard-disk-drive areal storage density.
Figure 1. Exponential growth in areal storage density has been characteristic of the hard-disk-drive industry since its inception in 1957. Inflection points in the compound growth rate (CGR) are coincident with the introduction of magnetoresistive (MR) read heads in 1992 and giant magnetoresistive (GMR) read heads in 1998. As a result of such innovations, the growth of the information economy and the increasing economic competition have been accompanied by an increase in the CGR. Today, progress is being sustained by the introduction of antiferromagnetic coupled media. As yet, there is no sign that the CGR is slowing, but some reduction is expected (dashed line) as areal densities approach physical limits imposed by the thermal instability of very small magnetic domains. The research history of magnetic read heads is shown in purple.
Figure 1. Exponential growth in areal storage density has been characteristic of the hard-disk-drive industry since its inception in 1957. Inflection points in the compound growth rate (CGR) are coincident with the introduction of magnetoresistive (MR) read heads in 1992 and giant magnetoresistive (GMR) read heads in 1998. As a result of such innovations, the growth of the information economy and the increasing economic competition have been accompanied by an increase in the CGR. Today, progress is being sustained by the introduction of antiferromagnetic coupled media. As yet, there is no sign that the CGR is slowing, but some reduction is expected (dashed line) as areal densities approach physical limits imposed by the thermal instability of very small magnetic domains. The research history of magnetic read heads is shown in purple.
In such a dynamic industry, time to market is everything. The company that first markets a new product garners most of the profits. Conversely, a new technology that is not in the marketplace rapidly loses its value. Selling old IT hardware is like selling old fish.
The importance of innovation
The miniaturization of IT components is widely understood to be an important driver of the long-term price and performance trends just discussed. The physical laws that guide the optimum scaling of transistors 1 are relatively simple, but if the shrinking of devices and circuits were straightforward, companies like ours could just spend their money on product development—and perhaps a little research with the sort of short time horizon that is easily justified to stockholders. But in fact, continued progress in miniaturization requires the profound innovation that often flows from highly exploratory or fundamental research.
All manufacturing tools and processes must constantly be improved and reinvented to address shrinking dimensions and tighter tolerances. Components that obey no simple scaling rules, such as wiring systems, must also be shrunk. And simple scaling rules inevitably break down as sizes get ever smaller. Eventually, quantum confinement effects and tunneling currents dominate the device design. To overcome constraints, new materials, including materials that do not yet exist, must be developed, tested, and understood, and then integrated into device structures. The need and opportunity for major innovations has been even greater in the magnetic-storage industry than in the microelectronics industry, perhaps because simple spatial scaling is thwarted by the laws of magnetism.
The importance of research
The incorporation of GMR sensors into hard-disk drives is an example of just how quickly a new and unexpected scientific discovery can energize an entire industry. When first observed 2 in 1988, the effect was significant only at cryogenic temperatures, and the expensive process of molecular beam epitaxy was needed to grow the layered metal structures with atomic precision. Stuart Parkin, an outstanding experimental physicist at IBM Corp who was intimately acquainted with the technological challenges of magnetic recording, saw a potential new technology. He also saw that low-cost sputter deposition techniques could be used to rapidly explore the enormous combinatorial space of possible layered magnetic structures and materials. Parkin and many others around the world explicated the underlying physics and quickly produced practical room-temperature GMR magnetic sensors. Teams of scientists and engineers then converged to uncover and solve myriad problems of materials compatibility, device reliability, manufacturing process control, and so on.
The first commercial GMR devices were magnetic field sensors sold in 1995. In 1997, IBM introduced GMR read heads to hard-disk drives (see figure 2), and all competing manufacturers followed as quickly as they could. Thus, a collective quantum behavior, unknown some 15 years ago, has become an integral part of the tens of millions of computers manufactured over the past year.
Figure 2. Giant magnetoresistance devices first became commercially available less than 10 years ago, but they are ubiquitous in today’s computers. The photo here shows a hard-disk-drive GMR read head that was introduced by IBM in 1997.
Figure 2. Giant magnetoresistance devices first became commercially available less than 10 years ago, but they are ubiquitous in today’s computers. The photo here shows a hard-disk-drive GMR read head that was introduced by IBM in 1997.
Here is another example. At IBM in the early 1980s, Bernard Meyerson and other scientists pursued fundamental studies of the gas-phase chemistry of organic compounds containing Si. Meyerson saw potential benefits for the epitaxial growth of Si crystals in a previously unexplored low-temperature, low-pressure regime. His subsequent invention of ultrahigh-vacuum Si epitaxy led to the fabrication of electronic devices that set a string of records for high-frequency performance. Scientists and engineers worked together to solve the development and manufacturing problems, and enabled IBM’s microelectronics division to offer new products for communications and to enter new markets. Today, despite a depressed semiconductor industry, we are seeing rapid growth of commercial applications for devices produced using Meyerson’s methods, including devices based on silicon—germanium electronics.
For every big win like the preceding examples, we could cite many more of lesser scale, and success stories like these are not limited to IBM or to the IT industry. At Lucent Technologies’ Bell Laboratories, for instance, fundamental studies of nonlinear optical processes led to the invention of optical fibers engineered for greatly reduced chromatic dispersion. Introduced to the market in 1994, Lucent’s True-Wave® optical fiber has become an industry standard for multiplex data transmission involving multiple wavelengths simultaneously carried on a single fiber. Similarly, studies at Bell Labs of the optical properties of rare-earth ions in glass hosts led to high-power erbium-doped fiber amplifiers. And studies of soliton dynamics led to pulse-shaping techniques for transmitting data, without repeaters, over very long distances.
Imperatives for success
To be first to the market with innovative products, an organization need not always make the initial scientific discovery. It must, however, have deep knowledge of the frontiers of science. The only way to foster that knowledge is to participate in the research endeavor. At the same time, knowledge of science is of little commercial value unless it can be related to business needs. Thus, the scientists who are engaged in research must also be familiar with the needs and problems of product development. The preceding considerations lead to two imperatives for success:
-
▸ To be a great industrial research institution, an organization must do product development.
-
▸ To be fast in bringing new technologies to market, an organization must do long-term exploratory research.
The first imperative is an obvious consequence of the need for speed. If research doesn’t bring a stream of new products and product innovations to the marketplace, it will not provide lasting value to the company. Furthermore, the transfer from a research organization to a development lab is a slow process. To speed that process, the inventors of new technology need to understand and be part of the subsequent product development.
The past two decades have not been kind to industrial research organizations that have not met the first imperative. In contrast, successful organizations, including IBM Research, remain strong in that respect. Some analysts have concluded that corporate investments in basic research must be declining. Actually, for successful companies that is not the case. For example, IBM’s research enterprise is still growing, but not faster than product development. We have built the channels needed for our innovations to flow into the market.
One might think that successful technology companies can forego long-term research and get around the second imperative—exploratory research—perhaps by purchasing small startup companies to ensure a constant influx of new ideas and products. However, that is generally not the case. Such companies may be successful, but their success is based not on technology leadership but on some other competency, such as low-cost manufacturing or distribution. Companies that strive to use technology to distinguish their products for consumers often invest in long-term research. For example, Microsoft Corp has made notable research investments in recent years, including hiring some excellent mathematical physicists. Intel Corp is investing in a series of research “lablets” located in close proximity to some top universities. The trend is clear, and we believe the investments are prudent. Any IT company that wants to be in business for the long haul will have to reinvent itself over and over, and would be wise to devote some resources to long-term exploratory research.
IBM is a case study in reinvention: In the 1930s, one of its products was meat scales, and in the 1950s, typewriters were a major product. So what long-term investment is IBM making in the physical sciences? Much of our exploratory research is aimed at processing, storing, and communicating information. For that, we are investigating new materials and devices, along with new architectures and algorithms. The physics of information is a unifying theme, guiding the long-term view of what is possible in IT.
Nanoscale science and engineering
We share some bemusement with the larger physics community at the sudden respectability of the word “nanotechnology.” After all, condensed matter physicists have been studying the physics of nanostructures for decades. What’s new is that manufactured devices now have dimensions that are measured in nanometers. In 2003, every company that wants to be a technology leader in microelectronics will have to use patterned lithography down to dimensions of about 90 nm, and use other techniques such as deposition or etching to manufacture yet smaller features. The “gate length” and other key dimensions of a transistor will be well below 50 nm. The physics of quantum confinement in structures of reduced dimensionality, pioneered in the late 1960s, is very relevant to the function of such nanodevices.
Progress in magnetic recording should continue for a good 10 years, with areal data storage densities reaching somewhere between 10 and 100 times today’s best values. The shrinking of Si transistors will probably continue for at least another 10 years, until the fairly hard physical limit of quantum tunneling is reached at gate lengths somewhere around 10–20 nm.
However, we see no physical or economic reason why those milestones should mark the end of the road for IT. Indeed, much of the excitement (and some unwarranted hype) surrounding nanotechnology is based on the promise of new devices operating at the molecular scale to sense, process, store, and communicate information. That promise is why companies are exploring the physics of electron transport in carbon nanotubes, 3 semiconductor quantum wires, 4 and a variety of organic molecules; it is also why they are fabricating primitive devices and circuits based on those new components. One such example is shown in figure 3. (See also the article by Jim Heath and Mark Ratner in Physics Today, May 2003, page 43.) The promise of molecular-scale devices has inspired companies to explore the physics of magnetic tunnel junctions and ionic transport in polymers and to use the knowledge they gain to develop new nonvolatile memory devices. Within IBM, we have responded to the promise by using our experience with scanning probe microscopy 5 to build Millipede, a novel scanning-probe approach to information storage. 6 (See Physics Today, October 2002, page 14.)
Figure 3. The carbon nanotube transistor shows potential to someday compete with the silicon transistor as a fast logical switch. (a) This drawing schematically illustrates an experimental carbon nanotube transistor. A potential applied to the overlying gate electrode controls the injection of charge carriers from the source or drain electrode into the nanotube. Although the arrangement shown is similar to that of an Si field-effect transistor, the recently elucidated physics of charge injection and transport (see ref. 4) is quite different. (b) The device characteristics show that a single nanotube, approximately 1.4 nm in diameter and gated by an electrical potential V g on the order of 1 volt, is capable of modulating several microamperes of current. Such modulation is remarkable for a switching device with so narrow a conducting channel. It would be comparable to that of the best exploratory Si transistors if they could be scaled to similarly narrow widths. Further improvements in the contact metallurgy and geometry of the device may result in a switch with impressive high-frequency performance. Nevertheless, the applications of carbon nanotube transistors will be limited unless practical manufacturing processes can be developed. Many laboratories are currently pursuing the improved synthetic paths and methods of chemical self-assembly that will be necessary.
Figure 3. The carbon nanotube transistor shows potential to someday compete with the silicon transistor as a fast logical switch. (a) This drawing schematically illustrates an experimental carbon nanotube transistor. A potential applied to the overlying gate electrode controls the injection of charge carriers from the source or drain electrode into the nanotube. Although the arrangement shown is similar to that of an Si field-effect transistor, the recently elucidated physics of charge injection and transport (see ref. 4) is quite different. (b) The device characteristics show that a single nanotube, approximately 1.4 nm in diameter and gated by an electrical potential V g on the order of 1 volt, is capable of modulating several microamperes of current. Such modulation is remarkable for a switching device with so narrow a conducting channel. It would be comparable to that of the best exploratory Si transistors if they could be scaled to similarly narrow widths. Further improvements in the contact metallurgy and geometry of the device may result in a switch with impressive high-frequency performance. Nevertheless, the applications of carbon nanotube transistors will be limited unless practical manufacturing processes can be developed. Many laboratories are currently pursuing the improved synthetic paths and methods of chemical self-assembly that will be necessary.
The problem of nanomanufacturing
To build a few very small devices is relatively easy. To manufacture products containing hundreds of millions of small interconnected devices is the challenge. The problem of nanomanufacturing is really a problem in information processing. Today, many gigabytes of compressed digital data are needed to define the circuit and device patterns of a typical microprocessor. The manufacturing process must transfer that information, at a very low error rate, from a hard-disk drive in the design center to pixels on a set of lithographic masks, and then to intricate patterns of metals, insulators, and impurities embedded in a Si wafer. As the complexity, and thus the information content, of the manufactured product increases, the economically acceptable error rate decreases.
Transferring information without introducing unacceptable error has historically been achieved largely by manufacturing processes and tools of ever greater precision and reliability. But on fundamental grounds, errors can never be totally eliminated from a manufacturing process. Increasingly, the challenge to scientists and engineers is to implement effective and affordable error-correction processes. The correction of photolithographic mask errors provides an example. Stray contaminants can cause errors in the pixels of a mask, and even a single pixel error can result in a flaw in every chip patterned from the imperfect mask. Outlining the methods used to find and correct such errors is beyond the scope of this article, but we note that repair of mask defects by femtosecond laser machining is the latest advance. Stemming from exploratory research into ultrafast nonlinear optical processes, the technique has, in the past few years, been turned into an essential and standard process for lithographic mask fabrication at IBM. 7
As device dimensions shrink toward the molecular scale, the cost of low-error-rate lithographic patterning and various error correction processes will continue to grow. That trend will generate increased opportunities and economic incentives to exploit processes of natural pattern formation or self-assembly to reduce the costs of building complex systems. Physicists have developed some profound ideas about the emergence of complexity in dissipative systems—renormalization and hierarchy of energy scales and self-organized criticality, to name just a few. There is more to be done, both in the development of fundamental concepts and in the application of those concepts to practical problems of nanomanufacturing.
For inspiration, increasing numbers of physicists are looking to the world of biology. Indeed, some scientists have speculated that the density of classical information in biological systems must be close to the physical limit. We believe the systems of IT will eventually approach similar limits. However, contrary to some of the hype surrounding nanotechnology, simple processes of self-assembly, in which the starting components are allowed to equilibrate, can never produce complex, hierarchically organized structures. The error rates will be too high. Entropy will win.
Some source of low-error-rate information is necessary to break symmetries, nucleate a variety of component structures, establish length scales, and guide the assembly process. In living systems, DNA is the source. In microelectronics manufacturing, lithography plays the analogous role. The stunning thing about biological manufacturing processes is that they use digital information so sparingly. Less than a gigabyte of digital information in DNA guides the assembly of something as complex as a human, but many gigabytes are required to guide the industrial assembly of something as comparatively simple as a microprocessor.
Thus, we at IBM are exploring a variety of self-assembly processes for IT with the goal of reducing (but not eliminating) the dependency on expensive lithographic manufacturing processes. We participate in the worldwide effort to synthesize new nanostructured materials and explore the potential application of those materials to IT hardware. 8,9 We and others are continuing to invent new instruments and to advance the capabilities of existing ones for nanoscale imaging and the characterization of materials and device structures. Figure 4, for example, shows how a low-energy electron microscope can track the growth of a one-monolayer crystal of pentacene, an organic semiconductor.
Figure 4. The growth of a single monolayer crystal of pentacene can be followed in real time with a low-energy electron microscope (LEEM). Molecules of pentacene, an organic semiconductor, are vacuum deposited on a crystalline silicon surface and agglomerate into a dendritic crystal by diffusion-limited aggregation (DLA). Images (a) through (c) show the nucleation and growth of the first molecular layer; (d) through (f) show nucleation and growth of the second and third layers. Scientists strive to create conditions that result in monolayers with a large lateral dimension (about 50 µm here) because the elimination of numerous grain boundaries results in organic semiconducting films with higher carrier mobility. In general, a basic understanding of the mechanisms of crystal nucleation and surface diffusion is extremely valuable in guiding the experimental fabrication of organic and molecular electronic devices. (g) The fractal dimension of single-monolayer pentacene islands depends on the degree of coverage. But, independent of the experimental conditions, including growth rate variations and substrate surface preparation as illustrated by the different colors and symbols, the fractal dimension follows a universal curve once the coverage exceeds about 0.05 monolayer.
Figure 4. The growth of a single monolayer crystal of pentacene can be followed in real time with a low-energy electron microscope (LEEM). Molecules of pentacene, an organic semiconductor, are vacuum deposited on a crystalline silicon surface and agglomerate into a dendritic crystal by diffusion-limited aggregation (DLA). Images (a) through (c) show the nucleation and growth of the first molecular layer; (d) through (f) show nucleation and growth of the second and third layers. Scientists strive to create conditions that result in monolayers with a large lateral dimension (about 50 µm here) because the elimination of numerous grain boundaries results in organic semiconducting films with higher carrier mobility. In general, a basic understanding of the mechanisms of crystal nucleation and surface diffusion is extremely valuable in guiding the experimental fabrication of organic and molecular electronic devices. (g) The fractal dimension of single-monolayer pentacene islands depends on the degree of coverage. But, independent of the experimental conditions, including growth rate variations and substrate surface preparation as illustrated by the different colors and symbols, the fractal dimension follows a universal curve once the coverage exceeds about 0.05 monolayer.
No one can be sure that carbon nanotube transistors or scanning-probe storage will dominate the future of IT. Surely, though, actively exploring the physical principles of such devices will create greater awareness and agility in responding to new developments from anywhere in the world.
Physics, biology, and computer science
IBM has not, traditionally, focused its research on biology. Nonetheless, IBM innovations aimed initially at IT have had significant impacts on biology and medicine. For example, laser ablation techniques that were originally targeted at patterning modules for chip interconnections have revolutionized corrective-vision eye surgery. 10 A more recent example is the direct mechanical transduction of biomolecular recognition in DNA hybridization using micro-machined cantilevers. 11
What has really sparked interest at IBM, however, is the rapid growth of sales of IT to customers in the life sciences industry. We have found that there is great value in understanding the scientific problems faced by those customers, and clearly there is great potential to discover exciting science at the interface of physics, biology, and computer science. Thus we are pursuing a variety of projects in computational biology and bioinformatics, with physicists often playing a starring role.
An example of our research in bioinformatics is the Genes@Work software package. The application 12 analyzes patterns in gene expression data (see figure 5) obtained from microarrays built by lithographic processes. The software is based on insights from statistical physics regarding the extraction of subtle patterns from large and noisy data matrices. Genes@Work has been particularly useful for extracting distinct patterns of gene expression associated with a given pathology, such as a specific type of cancer, and has already helped to generate exciting results in biology. 13 Such successes have stimulated further research into the sources of noise in these microarrays. 14
Figure 5. Gene expression analysis is a difficult problem in pattern recognition. IBM’s Genes@Work schema is one approach to solving part of that problem. (a) A population sharing the same disease—for example, a certain type of cancer—may be characterized by a gene expression matrix comprising expression values (colors) for a large set of genes (columns) and measured for different cells (rows) that share the same disease. One of many goals of gene expression analysis is to find gene expression profiles for the specific disease. In a typical experiment, each row in the gene expression matrix would contain measured expression values for more than 10 000 genes. Only a small subset of those genes is involved in a given disease, and most likely only a subset of cells falls into any particular gene expression profile. So the patterns of interest are simply defined as a submatrix of the whole gene expression matrix for which all the rows are “similar” in a statistically meaningful sense. (In the very simplified schematic shown here, the submatrix rows are, in fact, identical.) Because of the inherent noise in the measurement of the expression values and the large number of irrelevant genes, statistical analysis is crucial. The detailed analysis steps include statistical noise analysis of the gene expression data, novel gene-specific normalization, pattern discovery using an efficient search algorithm, and evaluation of statistical relevance for the discovered patterns.
Figure 5. Gene expression analysis is a difficult problem in pattern recognition. IBM’s Genes@Work schema is one approach to solving part of that problem. (a) A population sharing the same disease—for example, a certain type of cancer—may be characterized by a gene expression matrix comprising expression values (colors) for a large set of genes (columns) and measured for different cells (rows) that share the same disease. One of many goals of gene expression analysis is to find gene expression profiles for the specific disease. In a typical experiment, each row in the gene expression matrix would contain measured expression values for more than 10 000 genes. Only a small subset of those genes is involved in a given disease, and most likely only a subset of cells falls into any particular gene expression profile. So the patterns of interest are simply defined as a submatrix of the whole gene expression matrix for which all the rows are “similar” in a statistically meaningful sense. (In the very simplified schematic shown here, the submatrix rows are, in fact, identical.) Because of the inherent noise in the measurement of the expression values and the large number of irrelevant genes, statistical analysis is crucial. The detailed analysis steps include statistical noise analysis of the gene expression data, novel gene-specific normalization, pattern discovery using an efficient search algorithm, and evaluation of statistical relevance for the discovered patterns.
Algorithms that extract patterns from complex data are the basis of many of the products and services supplied by the IT industry. Yet such algorithms are generally inferior to those implemented in our own brains, and fundamental breakthroughs in pattern discovery are rare. Pattern-recognition research is linked to information theory, control theory, statistical physics, dynamical systems theory, and mathematical optimization theory.
As IT becomes more complex, with individual components tied to ever more complex networks, emergent behavior promises to become a feature, perhaps a problematic one. An interesting example of such behavior became evident in a virtual simulation designed to study the dynamics of future information economies. IBM developed a software platform that supported the interaction of tens of thousands of economically motivated software agents or “shopbots,” each exercising simple pricing algorithms and negotiation protocols. In a realistic simulation of an auction, the shopbots consistently obtained greater economic gains from trade than their human counterparts. The simulation suggests that, in the future, people may increasingly entrust software agents with important economic decisions, so long as that does not introduce undue risk. The collective dynamical behavior of the shopbots was difficult to predict, except by simulation, and it depended on the interaction rules in nonintuitive ways. The ideas and tools of nonlinear dynamics were useful aids to understanding the observations. 15
Quantum information
The concept of information has been transformed into a measurable, rigorously defined construct of physics, as fundamental as entropy. Information theory, originally limited to the realm of classical physics, is now a quantum theory, 16 and classical information is seen as a subset of quantum information, as shown in figure 6. The ability to reliably manipulate quantum information would enable quantum computing and other feats of information processing and communication that cannot be duplicated in the classical world. (See the Physics Today articles by Daniel Gottesman and Hoi-Kwong Lo, November 2000, page 22, and Barbara Terhal, Michael Wolf, and Andrew Doherty, April 2003, page 46.)
Figure 6. Quantum information subsumes classical information. At present, information technology is essentially limited to the processing, storage, and communication of classical information, but future information technologies may not be so limited. Quantum information allows operations that cannot be performed in a world of classical information.
Figure 6. Quantum information subsumes classical information. At present, information technology is essentially limited to the processing, storage, and communication of classical information, but future information technologies may not be so limited. Quantum information allows operations that cannot be performed in a world of classical information.
Of course, assembling a physical system that maintains quantum coherence for usefully long times remains a difficult technical problem. Building a general-purpose quantum computer introduces the more daunting problem of how to steer the dynamical evolution of a quantum coherent system without introducing significant probabilities for transitions to undesirable error states. Nevertheless, experiments at various laboratories have clearly demonstrated the potential to perform at least a few successive logical operations.
What will be the future applications of quantum information, and when will commercial products hit the market? We don’t know, although commercial quantum cryptographic systems, in which information is conveyed via the polarization states of single photons, appear to be approaching practicality. The ability to perform even a few reliable quantum logical operations—that is, logical operations performed by quantum-coherent physical systems—should enable useful new capabilities in areas like communications, metrology, and process feedback control. We are confident that scientists who are deeply familiar with quantum information will generate useful insights into classical IT.
In the past few decades, IT has grown rapidly to become a major component of the economy 17 With that growth has come a great broadening of the very meaning of IT. The ongoing “informatization” of biology, and the gathering momentum of the biotech revolution suggest that the broadening will continue. The past few decades have also seen information itself move to the center of physics. As a result, physicists are in an ideal position to help lead the emergence of the broad and truly ubiquitous information economy that so many have predicted.
REFERENCES
Tom Theis (ttheis@us.ibm.com) is the director of physical sciences at IBM Corp’s Thomas J. Watson Research Center in York-town Heights, New York. Paul Horn (horn@watson.ibm.com) is the center’s senior vice president for research.
