Steve Blank

Artificial Intelligence is a once-in-a lifetime commercial and defense game changer

(download a PDF of this article here)

Hundreds of billions in public and private capital is being invested in Artificial Intelligence (AI) and Machine Learning companies. The number of patents filed in 2021 is more than 30 times higher than in 2015 as companies and countries across the world have realized that AI and Machine Learning will be a major disruptor and potentially change the balance of military power.

Until recently, the hype exceeded reality. Today, however, advances in AI in several important areas (here, here, here, here and here) equal and even surpass human capabilities.

If you haven’t paid attention, now’s the time.

Artificial Intelligence and the Department of Defense (DoD)
The Department of Defense has thought that Artificial Intelligence is such a foundational set of technologies that they started a dedicated organization- the JAIC – to enable and implement artificial intelligence across the Department. They provide the infrastructure, tools, and technical expertise for DoD users to successfully build and deploy their AI-accelerated projects.

Some specific defense related AI applications are listed later in this document.

We’re in the Middle of a Revolution
Imagine it’s 1950, and you’re a visitor who traveled back in time from today. Your job is to explain the impact computers will have on business, defense and society to people who are using manual calculators and slide rules. You succeed in convincing one company and a government to adopt computers and learn to code much faster than their competitors /adversaries. And they figure out how they could digitally enable their business – supply chain, customer interactions, etc. Think about the competitive edge they’d have by today in business or as a nation. They’d steamroll everyone.

That’s where we are today with Artificial Intelligence and Machine Learning. These technologies will transform businesses and government agencies. Today, 100s of billions of dollars in private capital have been invested in 1,000s of AI startups. The U.S. Department of Defense has created a dedicated organization to ensure its deployment.

But What Is It?
Compared to the classic computing we’ve had for the last 75 years, AI has led to new types of applications, e.g. facial recognition; new types of algorithms, e.g. machine learning; new types of computer architectures, e.g. neural nets; new hardware, e.g. GPUs; new types of software developers, e.g. data scientists; all under the overarching theme of artificial intelligence. The sum of these feels like buzzword bingo. But they herald a sea change in what computers are capable of doing, how they do it, and what hardware and software is needed to do it.

This brief will attempt to describe all of it.

New Words to Define Old Things
One of the reasons the world of AI/ML is confusing is that it’s created its own language and vocabulary. It uses new words to define programming steps, job descriptions, development tools, etc. But once you understand how the new world maps onto the classic computing world, it starts to make sense. So first a short list of some key definitions.

AI/ML – a shorthand for Artificial Intelligence/Machine Learning

Artificial Intelligence (AI) – a catchall term used to describe “Intelligent machines” which can solve problems, make/suggest decisions and perform tasks that have traditionally required humans to do. AI is not a single thing, but a constellation of different technologies.

Machine Learning (ML) – a subfield of artificial intelligence. Humans combine data with algorithms (see here for a list) to train a model using that data. This trained model can then make predications on new data (is this picture a cat, a dog or a person?) or decision-making processes (like understanding text and images) without being explicitly programmed to do so.

Machine learning algorithms – computer programs that adjust themselves to perform better as they are exposed to more data. The “learning” part of machine learning means these programs change how they process data over time. In other words, a machine-learning algorithm can adjust its own settings, given feedback on its previous performance in making predictions about a collection of data (images, text, etc.).

Deep Learning/Neural Nets – a subfield of machine learning. Neural networks make up the backbone of deep learning. (The “deep” in deep learning refers to the depth of layers in a neural network.) Neural nets are effective at a variety of tasks (e.g., image classification, speech recognition). A deep learning neural net algorithm is given massive volumes of data, and a task to perform – such as classification. The resulting model is capable of solving complex tasks such as recognizing objects within an image and translating speech in real time. In reality, the neural net is a logical concept that gets mapped onto a physical set of specialized processors. See here.)

Data Science – a new field of computer science. Broadly it encompasses data systems and processes aimed at maintaining data sets and deriving meaning out of them. In the context of AI, it’s the practice of people who are doing machine learning.

Data Scientists – responsible for extracting insights that help businesses make decisions. They explore and analyze data using machine learning platforms to create models about customers, processes, risks, or whatever they’re trying to predict.

What’s Different? Why is Machine Learning Possible Now?
To understand why AI/Machine Learning can do these things, let’s compare them to computers before AI came on the scene. (Warning – simplified examples below.)

Classic Computers

For the last 75 years computers (we’ll call these classic computers) have both shrunk to pocket size (iPhones) and grown to the size of warehouses (cloud data centers), yet they all continued to operate essentially the same way.

Classic Computers – Programming
Classic computers are designed to do anything a human explicitly tells them to do. People (programmers) write software code (programming) to develop applications, thinking a priori about all the rules, logic and knowledge that need to be built in to an application so that it can deliver a specific result. These rules are explicitly coded into a program using a software language (Python, JavaScript, C#, Rust, …).

Classic Computers –  Compiling
The code is then compiled using software to translate the programmer’s source code into a version that can be run on a target computer/browser/phone. For most of today’s programs, the computer used to develop and compile the code does not have to be that much faster than the one that will run it.

Classic Computers – Running/Executing Programs
Once a program is coded and compiled, it can be deployed and run (executed) on a desktop computer, phone, in a browser window, a data center cluster, in special hardware, etc. Programs/applications can be games, social media, office applications, missile guidance systems, bitcoin mining, or even operating systems e.g. Linux, Windows, IOS. These programs run on the same type of classic computer architectures they were programmed in.

Classic Computers – Software Updates, New Features
For programs written for classic computers, software developers receive bug reports, monitor for security breaches, and send out regular software updates that fix bugs, increase performance and at times add new features.

Classic Computers-  Hardware
The CPUs (Central Processing Units) that write and run these Classic Computer applications all have the same basic design (architecture). The CPUs are designed to handle a wide range of tasks quickly in a serial fashion. These CPUs range from Intel X86 chips, and the ARM cores on Apple M1 SoC, to the z15 in IBM mainframes.

Machine Learning

In contrast to programming on classic computing with fixed rules, machine learning is just like it sounds – we can train/teach a computer to “learn by example” by feeding it lots and lots of examples. (For images a rule of thumb is that a machine learning algorithm needs at least 5,000 labeled examples of each category in order to produce an AI model with decent performance.) Once it is trained, the computer runs on its own and can make predictions and/or complex decisions.

Just as traditional programming has three steps – first coding a program, next compiling it and then running it – machine learning also has three steps: training (teaching), pruning and inference (predicting by itself.)

Machine Learning – Training
Unlike programing classic computers with explicit rules, training is the process of “teaching” a computer to perform a task e.g. recognize faces, signals, understand text, etc. (Now you know why you’re asked to click on images of traffic lights, cross walks, stop signs, and buses or type the text of scanned image in ReCaptcha.) Humans provide massive volumes of “training data” (the more data, the better the model’s performance) and select the appropriate algorithm to find the best optimized outcome. (See the detailed “machine learning pipeline” section for the gory details.)

By running an algorithm selected by a data scientist on a set of training data, the Machine Learning system generates the rules embedded in a trained model. The system learns from examples (training data), rather than being explicitly programmed. (See the “Types of Machine Learning” section for more detail.) This self-correction is pretty cool. An input to a neural net results in a guess about what that input is. The neural net then takes its guess and compares it to a ground-truth about the data, effectively asking an expert “Did I get this right?” The difference between the network’s guess and the ground truth is its error. The network measures that error, and walks the error back over its model, adjusting weights to the extent that they contributed to the error.)

Just to make the point again: The algorithms combined with the training data – not external human computer programmers – create the rules that the AI uses. The resulting model is capable of solving complex tasks such as recognizing objects it’s never seen before, translating text or speech, or controlling a drone swarm.

(Instead of building a model from scratch you can now buy, for common machine learning tasks, pretrained models from others and here, much like chip designers buying IP Cores.)

Machine Learning Training – Hardware
Training a machine learning model is a very computationally intensive task. AI hardware must be able to perform thousands of multiplications and additions in a mathematical process called matrix multiplication. It requires specialized chips to run fast. (See the AI semiconductor section for details.)

Machine Learning – Simplification via pruning, quantization, distillation
Just like classic computer code needs to be compiled and optimized before it is deployed on its target hardware, the machine learning models are simplified and modified (pruned) to use less computing power, energy, and  memory before they’re deployed to run on their hardware.

Machine Learning – Inference Phase
Once the system has been trained it can be copied to other devices and run. And the computing hardware can now make inferences (predictions) on new data that the model has never seen before.

Inference can even occur locally on edge devices where physical devices meet the digital world (routers, sensors, IOT devices), close to the source of where the data is generated. This reduces network bandwidth issues and eliminates latency issues.

Machine Learning Inference – Hardware
Inference (running the model) requires substantially less compute power than training. But inference also benefits from specialized AI chips. (See the AI semiconductor section for details.)

Machine Learning – Performance Monitoring and Retraining
Just like classic computers where software developers do regular software updates to fix bugs and increase performance and add features, machine learning models also need to be updated regularly by adding new data to the old training pipelines and running them again. Why?

Over time machine learning models get stale. Their real-world performance generally degrades over time if they are not updated regularly with new training data that matches the changing state of the world. The models need to be monitored and retrained regularly for data and/or concept drift, harmful predictions, performance drops, etc. To stay up to date, the models need to re-learn the patterns by looking at the most recent data that better reflects reality.

One Last Thing – “Verifiability/Explainability”
Understanding how an AI works is essential to fostering trust and confidence in AI production models.

Neural Networks and Deep Learning differ from other types of Machine Learning algorithms in that they have low explainability. They can generate a prediction, but it is very difficult to understand or explain how it arrived at its prediction. This “explainability problem” is often described as a problem for all of AI, but it’s primarily a problem for Neural Networks and Deep Learning. Other types of Machine Learning algorithms – for example decision trees or linear regression– have very high explainability. The results of the five-year DARPA Explainable AI Program (XAI) are worth reading here.

So What Can Machine Learning Do?

It’s taken decades but as of today, on its simplest implementations, machine learning applications can do some tasks better and/or faster than humans. Machine Learning is most advanced and widely applied today in processing text (through Natural Language Processing) followed by understanding images and videos (through Computer Vision) and analytics and anomaly detection. For example:

Recognize and Understand Text/Natural Language Processing
AI is better than humans on basic reading comprehension benchmarks like SuperGLUE and SQuAD and their performance on complex linguistic tasks is almost there. Applications: GPT-3, M6, OPT-175B, Google Translate, Gmail Autocomplete, Chatbots, Text summarization.

Write Human-like Answers to Questions and Assist in Writing Computer Code
An AI can write original text that is indistinguishable from that created by humans. Examples GPT-3, Wu Dao 2.0 or generate computer code. Example GitHub Copilot, Wordtune

Recognize and Understand Images and video streams
An AI can see and understand what it sees. It can identify and detect an object or a feature in an image or video. It can even identify faces. It can scan news broadcasts or read and assess text that appears in videos. It has uses in threat detection –  airport security, banks, and sporting events. In medicine to interpret MRI’s or to design drugs. And in retail to scan and analyze in-store imagery to intuitively determine inventory movement. Examples of ImageNet benchmarks here and here

Turn 2D Images into 3D Rendered Scenes
AI using “NeRFs “neural radiance fields” can take 2d snapshots and render a finished 3D scene in realtime to create avatars or scenes for virtual worlds, to capture video conference participants and their environments in 3D, or to reconstruct scenes for 3D digital maps. The technology is an enabler of the metaverse, generating digital representations of real environments that creators can modify and build on. And self driving cars are using NeRF’s to render city-scale scenes spanning multiple blocks.

Detect Changes in Patterns/Recognize Anomalies
An AI can recognize patterns which don’t match the behaviors expected for a particular system, out of millions of different inputs or transactions. These applications can discover evidence of an attack on financial networks, fraud detection in insurance filings or credit card purchases; identify fake reviews; even tag sensor data in industrial facilities that mean there’s a safety issue. Examples here, here and here.

Power Recommendation Engines
An AI can provide recommendations based on user behaviors used in ecommerce to provide accurate suggestions of products to users for future purchases based on their shopping history. Examples: Netflix, TikTok, CrossingMinds and Recommendations AI

Recognize and Understand Your Voice
An AI can understand spoken language. Then it can comprehend what is being said and in what context. This can enable chatbots to have a conversation with people. It can record and transcribe meetings. (Some versions can even read lips to increase accuracy.) Applications: Siri/Alexa/Google Assistant. Example here

Create Artificial Images
AI can ​create artificial ​images​ (DeepFakes) that ​are​ indistinguishable ​from​ real ​ones using Generative Adversarial Networks.​ Useful in ​entertainment​, virtual worlds, gaming, fashion​ design, etc. Synthetic faces are now indistinguishable and more trustworthy than photos of real people. Paper here.

Create Artist Quality Illustrations from A Written Description
AI can generate images from text descriptions, creating anthropomorphized versions of animals and objects, combining unrelated concepts in plausible ways. An example application is Dall-E

Generative Design of Physical Products
Engineers can input design goals into AI-driven generative design software, along with parameters such as performance or spatial requirements, materials, manufacturing methods, and cost constraints. The software explores all the possible permutations of a solution, quickly generating design alternatives. Example here.

Sentiment Analysis
An AI leverages deep natural language processing, text analysis, and computational linguistics to gain insight into customer opinion, understanding of consumer sentiment, and measuring the impact of marketing strategies. Examples: Brand24, MonkeyLearn

AI in National Security

Much like the dual-use/dual-nature of classical computers AI developed for commercial applications can also be used for national security.

AI/ML and Ubiquitous Technical Surveillance
AI/ML have made most cities untenable for traditional tradecraft. Machine learning can integrate travel data (customs, airline, train, car rental, hotel, license plate readers…,) integrate feeds from CCTV cameras for facial recognition and gait recognition, breadcrumbs from wireless devices and then combine it with DNA sampling. The result is automated persistent surveillance.

China’s employment of AI as a tool of repression and surveillance of the Uyghurs is a reminder of a dystopian future of how totalitarian regimes will use AI-enabled ubiquitous surveillance to repress and monitor its own populace.

AI/ML on the Battlefield
AI will enable new levels of performance and autonomy for weapon systems. Autonomously collaborating assets (e.g., drone swarms, ground vehicles) that can coordinate attacks, ISR missions, & more.

Fusing and making sense of sensor data (detecting threats in optical /SAR imagery, classifying aircraft based on radar returns, searching for anomalies in radio frequency signatures, etc.) Machine learning is better and faster than humans in finding targets hidden in a high-clutter background. Automated target detection and fires from satellite/UAV.

For example, an Unmanned Aerial Vehicle (UAV) or Unmanned Ground Vehicles with on board AI edge computers could use deep learning to detect and locate concealed chemical, biological and explosive threats by fusing imaging sensors and chemical/biological sensors.

Other examples include:

Use AI/ML countermeasures against adversarial, low probability of intercept/low probability of detection (LPI/LPD) radar techniques in radar and communication systems.

Given sequences of observations of unknown radar waveforms from arbitrary emitters without a priori knowledge, use machine learning to develop behavioral models to enable inference of radar intent and threat level, and to enable prediction of future behaviors.

For objects in space, use machine learning to predict and characterize a spacecrafts possible actions, its subsequent trajectory, and what threats it can pose from along that trajectory. Predict the outcomes of finite burn, continuous thrust, and impulsive maneuvers.

AI empowers other applications such as:

• Flight Operations Planning Decision Aid Tool for Strike Operations Aboard Aircraft Carriers
• Automated Battle management – air and missile defense, army/navy tactical…

AI/ML in Collection
The front end of intelligence collection platforms has created a firehose of data that have overwhelmed human analysts. “Smart” sensors coupled with inference engines that can pre-process raw intelligence and prioritize what data to transmit and store –helpful in degraded or low-bandwidth environments.

Human-Machine Teaming in Signals Intelligence
Applications with embedded intelligence have already begun to appear in commercial applications thanks to massive language models. For example – Copilot as a pair-programmer in Microsoft Visual Studio VSCode. It’s not hard to imagine an AI that can detect and isolate anomalies and other patterns of interest in all sorts of signal data faster and more reliably than human operators.

AI-enabled natural language processing, computer vision, and audiovisual analysis can vastly reduce manual data processing. Advances in speech-to-text transcription and language analytics now enable reading comprehension, question answering, and automated summarization of large quantities of text. This not only prioritizes the work of human analysts, it’s a major force multiplier

AI can also be used to automate data conversion such as translations and decryptions, accelerating the ability to derive actionable insights.

Human-Machine Teaming in Tasking and Dissemination
AI-enabled systems will automate and optimize tasking and collection for platforms, sensors, and assets in near-real time in response to dynamic intelligence requirements or changes in the environment.

AI will be able to automatically generate machine-readable versions of intelligence products and disseminate them at machine speed so that computer systems across the IC and the military can ingest and use them in real time without manual intervention.

Human-Machine Teaming in Exploitation and Analytics
AI-enabled tools can augment filtering, flagging, and triage across multiple data sets. They can identify connections and correlations more efficiently and at a greater scale than human analysts, and can flag those findings and the most important content for human analysis.

AI can fuse data from multiple sources, types of intelligence, and classification levels to produce accurate predictive analysis in a way that is not currently possible. This can improve indications and warnings for military operations and active cyber defense.

AI/ML Information warfare
Nation states have used AI systems to enhance disinformation campaigns and cyberattacks. This included using “DeepFakes” (fake videos generated by a neural network that are nearly indistinguishable from reality). They are harvesting data on Americans to build profiles of our beliefs, behavior, and biological makeup for tailored attempts to manipulate or coerce individuals.

But because a large percentage of it is open-source AI is not limited to nation states, AI-powered cyber-attacks, deepfakes and AI software paired with commercially available drones can create “poor-man’s smart weapons” for use by rogue states, terrorists and criminals.

AI/ML Cyberwarfare
AI-enabled malware can learn and adapt to a system’s defensive measures, by probing a target system to look for system configuration and operational patterns and customize the attack payload to determine the most opportune time to execute the payload so to maximize the impact. Conversely, AI-enabled cyber-defensive tools can proactively locate and address network anomalies and system vulnerabilities.

Attacks Against AI – Adversarial AI
As AI proliferates, defeating adversaries will be predicated on defeating their AI and vice versa. As Neural Networks take over sensor processing and triage tasks, a human may only be alerted if the AI deems it suspicious. Therefore, we only need to defeat the AI to evade detection, not necessarily a human.

Adversarial attacks against AI fall into three types:

• Data misclassification– to generate false positive or negative results
• Synthetic data generation-to feed false information
• Data analysis – for AI-assisted classical attack generation

AI Attack Surfaces
Electronic Attack (EA), Electronic Protection (EP), Electronic Support (ES) all have analogues in the AI algorithmic domain. In the future, we may play the same game about the “Algorithmic Spectrum,” denying our adversaries their AI capabilities while defending ours. Other can steal or poison our models  or manipulate our training data.

We’re Just Getting Started
Progress in AI has been growing exponentially. The next 10 years will see a massive improvement on AI inference and training capabilities. This will require regular refreshes of the hardware– on the chip and cloud clusters – to take advantage. This is the AI version of Moore’s Law on steroids – applications that are completely infeasible today will be easy in 5 years.

What Can’t AI Do?

While AI can do a lot of things better than humans when focused on a narrow objective, there are many things it still can’t do. AI works well in specific domain where you have lots of data, time/resources to train, domain expertise to set the right goals/rewards during training, but that is not always the case.

For example AI models are only as good as the fidelity and quality of the training data. Having bad labels can wreak havoc on your training results. Protecting the integrity of the training data is critical.

In addition, AI is easily fooled by out-of-domain data (things it hasn’t seen before). This can happen by “overfitting” – when a model trains for too long on sample data or when the model is too complex, it can start to learn the “noise,” or irrelevant information, within the dataset. When the model memorizes the noise and fits too closely to the training set, the model becomes “overfitted,” and it is unable to generalize well to new data. If a model cannot generalize well to new data, then it will not be able to perform the classification or prediction tasks it was intended for. However, if you pause too early or exclude too many important features, you may encounter the opposite problem, and instead, you may “underfit” your model. Underfitting occurs when the model has not trained for enough time, or the input variables are not significant enough to determine a meaningful relationship between the input and output variables.

AI is also poor at estimating uncertainty /confidence (and explaining its decision-making). It can’t choose its own goals. (Executives need to define the decision that the AI will execute.  Without well-defined decisions to be made, data scientists will waste time, energy and money.) Except for simple cases an AI can’t (yet) figure out cause and effect or why something happened. It can’t think creatively or apply common sense.

AI is not very good at creating a strategy (unless it can pull from previous examples and mimic them, but then fails with the unexpected.) And it lacks generalized intelligence e.g. that can generalize knowledge and translate learning across domains.

All of these are research topics actively being worked on. Solving these will take a combination of high-performance computing, advanced AI/ML semiconductors, creative machine learning implementations and decision science. Some may be solved in the next decade, at least to a level where a human can’t tell the difference.

Where is AI and National Security Going Next?

In the near future AI may be able to predict the future actions an adversary could take and the actions a friendly force could take to counter these. The 20th century model loop of Observe–Orient–Decide and Act (OODA) is retrospective; an observation cannot be made until after the event has occurred. An AI-enabled decision-making cycle might be ‘sense–predict–agree–act’: AI senses the environment; predicts what the adversary might do and offers what a future friendly force response should be; the human part of the human–machine team agrees with this assessment; and AI acts by sending machine-to-machine instructions to the small, agile and many autonomous warfighting assets deployed en masse across the battlefield.

An example of this is DARPA’s ACE (Air Combat Evolution) program that is developing a warfighting concept for combined arms using a manned and unmanned systems. Humans will fight in close collaboration with autonomous weapon systems in complex environments with tactics informed by artificial intelligence.

A Once-in-a-Generation Event
Imagine it’s the 1980’s and you’re in charge of an intelligence agency. SIGINT and COMINT were analog and RF. You had worldwide collection systems with bespoke systems in space, air, underwater, etc. And you wake up to a world that shifts from copper to fiber. Most of your people, and equipment are going to be obsolete, and you need to learn how to capture those new bits. Almost every business processes needed to change, new organizations needed to be created, new skills were needed, and old ones were obsoleted. That’s what AI/ML is going to do to you and your agency.

The primary obstacle to innovation in national security is not technology, it is culture. The DoD and IC must overcome a host of institutional, bureaucratic, and policy challenges to adopting and integrating these new technologies. Many parts of our culture are resistant to change, reliant on traditional tradecraft and means of collection, and averse to risk-taking, (particularly acquiring and adopting new technologies and integrating outside information sources.)

History tells us that late adopters fall by the wayside as more agile and opportunistic governments master new technologies.

Carpe Diem.

Want more Detail?

Read on if you want to know about Machine Learning chips, see a sample Machine Learning Pipeline and learn about the four types of Machine Learning.

 

Details of a machine learning pipeline

This is a sample of the workflow (a pipeline) data scientists use to develop, deploy and maintain a machine learning model (see the detailed description here.)

(download a PDF of this article here)

Sources:

• Understanding AI Technology: Greg Allen, Chief of Strategy and Communications Joint Artificial Intelligence Center (JAIC), Department of Defense https://www.ai.mil/docs/Understanding%20AI%20Technology.pdf
• ·AI, Machine Learning, Deep Learning Explained Simply: Jun Wu https://towardsdatascience.com/ai-machine-learning-deep-learning-explained-simply-7b553da5b960
• The Democratization of Artificial Intelligence and Deep Learning: Databrickshttps://databricks.com/discover/pages/the-democratization-of-artificial-intelligence-and-deep-learning
• Final Report: National Security Report on Artificial Intelligencehttps://www.nscai.gov/wp-content/uploads/2021/03/Full-Report-Digital-1.pdf
• A Beginners Guide to Neural Nets and Deep Learning: Pathmind https://wiki.pathmind.com/neural-network
• IBM Cloud Learning: Overfiitting https://www.ibm.com/cloud/learn/overfitting

Filed under: Gordian Knot Center for National Security Innovation, Technology | 2 Comments »

If you think you understand quantum mechanics,
you don’t understand quantum mechanics

Richard Feynman

Tens of billions of public and private capital are being invested in Quantum technologies. Countries across the world have realized that quantum technologies can be a major disruptor of existing businesses and change the balance of military power. So much so, that they have collectively invested ~$24 billion in in quantum research and applications.

At the same time, a week doesn’t go by without another story about a quantum technology milestone or another quantum company getting funded. Quantum has moved out of the lab and is now the focus of commercial companies and investors. In 2021 venture capital funds invested over $2 billion in 90+ Quantum technology companies. Over a $1 billion of it going to Quantum computing companies. In the last six months quantum computing companies IonQ, D-Wave and Rigetti went public at valuations close to a billion and half dollars. Pretty amazing for computers that won’t be any better than existing systems for at least another decade – or more.  So why the excitement about quantum?

The Quantum Market Opportunity

While most of the IPOs have been in Quantum Computing, Quantum technologies are used in three very different and distinct markets: Quantum Computing, Quantum Communications and Quantum Sensing and Metrology.

All of three of these markets have the potential for being disruptive. In time Quantum computing could obsolete existing cryptography systems, but viable commercial applications are still speculative. Quantum communications could allow secure networking but are not a viable near-term business. Quantum sensors could create new types of medical devices, as well as new classes of military applications, but are still far from a scalable business.

It’s a pretty safe bet that 1) the largest commercial applications of quantum technologies won’t be the ones these companies currently think they’re going to be, and 2) defense applications using quantum technologies will come first. 3) if and when they do show up they’ll destroy existing businesses and create new ones.

We’ll describe each of these market segments in detail. But first a description of some quantum concepts.

Types of Quantum Computers

Quantum computers could potentially do things at speeds current computers cannot. Think of the difference of how fast you can count on your fingers versus how fast today’s computers can count. That’s the same order of magnitude speed-up a quantum computer could have over today’s computers for certain applications.

Quantum computers fall into four categories:

1. Quantum Emulator/Simulator
2. Quantum Annealer
3. NISQ – Noisy Intermediate Scale Quantum
4. Universal Quantum Computer – which can be a Cryptographically Relevant Quantum Computer (CRQC)

When you remove all the marketing hype, the only type that matters is #4 – a Universal Quantum Computer. And we’re at least a decade or more away from having those.

Quantum Emulator/Simulator
These are classical computers that you can buy today that simulate quantum algorithms. They make it easy to test and debug a quantum algorithm that someday may be able to run on a Universal Quantum Computer. Since they don’t use any quantum hardware they are no faster than standard computers.

Quantum Annealer is a special purpose quantum computer designed to only run combinatorial optimization problems, not general-purpose computing, or cryptography problems. D-Wave has defined and owned this space. While they have more physical Qubits than any other current system they are not organized as gate-based logical qubits. Currently this is a nascent commercial technology in search of a future viable market.

Noisy Intermediate-Scale Quantum (NISQ) computers. Think of these as prototypes of a Universal Quantum Computer – with several orders of magnitude fewer bits. (They currently have 50-100 qubits, limited gate depths, and short coherence times.) As they are short several orders of magnitude of Qubits, NISQ computers cannot perform any useful computation, however they are a necessary phase in the learning, especially to drive total system and software learning in parallel to the hardware development. Think of them as the training wheels for future universal quantum computers.

Universal Quantum Computers / Cryptographically Relevant Quantum Computers (CRQC)
This is the ultimate goal. If you could build a universal quantum computer with fault tolerance (i.e. millions of error corrected physical qubits resulting in thousands of logical Qubits), you could run quantum algorithms in cryptography, search and optimization, quantum systems simulations, and linear equations solvers. (See here for a list of hundreds quantum algorithms.) These all would dramatically outperform classical computation on large complex problems that grow exponentially as more variables are considered. Classical computers can’t attack these problems in reasonable times without so many approximations that the result is useless. We simply run out of time and transistors with classical computing on these problems. These special algorithms are what make quantum computers potentially valuable. For example, Grover’s algorithm solves the problem for the unstructured search of data. Further, quantum computers are very good at minimization / optimizations…think optimizing complex supply chains, energy states to form complex molecules, financial models, etc.

However, while all of these algorithms might have commercial potential one day, no one has yet to come up with a use for them that would radically transform any business or military application. Except for one – and that one keeps people awake at night.

It’s Shor’s algorithm for integer factorization – an algorithm that underlies much of existing public cryptography systems.

The security of today’s public key cryptography systems rests on the assumption that breaking into those with a thousand or more digits is practically impossible. It requires factoring into large prime numbers (e.g., RSA) or elliptic curve (e.g., ECDSA, ECDH) or finite fields (DSA) that can’t be done with any type of classic computer regardless of how large. Shor’s factorization algorithm can crack these codes if run on a Universal Quantum Computer. Uh-oh!

Estimates of when you can actually buy a fully error-corrected quantum computers vary from “never” to somewhere between 8 to 20 years from now. (Some optimists believe even earlier.)

Quantum Communication

Quantum communications ≠ quantum computers. A quantum network’s value comes from its ability to distribute entanglement. These communication devices manipulate the quantum properties of photons/particles of light to build Quantum Networks.

This market includes secure quantum key distribution, clock synchronization, random number generation and networking of quantum military sensors, computers, and other systems.

Quantum Cryptography/Quantum Key Distribution
Quantum Cryptography/Quantum Key Distribution can distribute keys between authorized partners connected by a quantum channel and a classical authenticated channel. It can be implemented via fiber optics or free space transmission. China transmitted entangled photons (at one pair of entangled particles per second) over 1,200 km in a satellite link, using the Micius satellite.

The Good: it can detect the presence of an eavesdropper, a feature not provided in standard cryptography. The Bad: Quantum Key Distribution can’t be implemented in software or as a service on a network and cannot be easily integrated into existing network equipment. It lacks flexibility for upgrades or security patches. Securing and validating Quantum Key Distribution is hard and it’s only one part of a cryptographic system.

The view from the National Security Agency (NSA) is that quantum-resistant (or post-quantum) cryptography is a more cost effective and easily maintained solution than quantum key distribution. NSA does not support the usage of QKD or QC to protect communications in National Security Systems. (See here.) They do not anticipate certifying or approving any Quantum Cryptography/Quantum Key Distribution security products for usage by National Security System customers unless these limitations are overcome. However, if you’re a commercial company these systems may be worth exploring.

Quantum Random Number Generators (GRGs)
Commercial Quantum Random Number Generators that use quantum effects (entanglement) to generate nondeterministic randomness are available today. (Government agencies can already make quality random numbers and don’t need these devices.)

Random number generators will remain secure even when a Cryptographically Relevant Quantum Computer is built.

Quantum Sensing and Metrology

Quantum sensors ≠ Quantum computers.

This segment consists of Quantum Sensing (quantum magnetometers, gravimeters, …), Quantum Timing (precise time measurement and distribution), and Quantum Imaging (quantum radar, low-SNR imaging, …) Each of these areas can create entirely new commercial products or entire new industries e.g. new classes of medical devices and military systems, e.g. anti-submarine warfare, detecting stealth aircraft, finding hidden tunnels and weapons of mass destruction. Some of these are achievable in the near term.

Quantum Timing
First-generation quantum timing devices already exist as microwave atomic clocks. They are used in GPS satellites to triangulate accurate positioning. The Internet and computer networks use network time servers and the NTP protocol to receive the atomic clock time from either the GPS system or a radio transmission.

The next generation of quantum clocks are even more accurate and use laser-cooled single ions confined together in an electromagnetic ion trap. This increased accuracy is not only important for scientists attempting to measure dark matter and gravitational waves, but miniaturized/ more accurate atomic clocks will allow precision navigation in GPS- degraded/denied areas, e.g. in commercial and military aircraft, in tunnels and caves, etc.

Quantum Imaging
Quantum imaging is one of the most interesting and near-term applications. First generation magnetometers such as superconducting quantum interference devices (SQUIDs) already exist. New quantum sensor types of imaging devices use entangled light, accelerometers, magnetometers, electrometers, gravity sensors. These allow measurements of frequency, acceleration, rotation rates, electric and magnetic fields, photons, or temperature with levels of extreme sensitivity and accuracy.

These new sensors use a variety of quantum effects: electronic, magnetic, or vibrational states or spin qubits, neutral atoms, or trapped ions. Or they use quantum coherence to measure a physical quantity. Or use quantum entanglement to improve the sensitivity or precision of a measurement, beyond what is possible classically.

Quantum Imaging applications can have immediate uses in archeology,  and profound military applications. For example, submarine detection using quantum magnetometers or satellite gravimeters could make the ocean transparent. It would compromise the survivability of sea-based nuclear deterrent by detecting and tracking subs deep underwater.

Quantum sensors and quantum radar from companies like Rydberg can be game changers.

Gravimeters or quantum magnetometers could also detect concealed tunnels, bunkers, and nuclear materials. Magnetic resonance imaging could remotely ID chemical and biological agents. Quantum radar or LIDAR would enable extreme detection of electromagnetic emissions, enhancing ELINT and electronic warfare capabilities. It can use fewer emissions to get the same detection result, for better detection accuracy at the same power levels – even detecting stealth aircraft.

Finally, Ghost imaging uses the quantum properties of light to detect distant objects using very weak illumination beams that are difficult for the imaged target to detect. It can increase the accuracy and lessen the amount of radiation exposed to a patient during x-rays. It can see through smoke and clouds. Quantum illumination is similar to ghost imaging but could provide an even greater sensitivity.

National and Commercial Efforts
Countries across the world are making major investments ~$24 billion in 2021 – in quantum research and applications.

Lessons Learned

• Quantum technologies are emerging and disruptive to companies and defense
• Quantum technologies cover Quantum Computing, Quantum Communications and Quantum Sensing and Metrology
  • Quantum computing could obsolete existing cryptography systems
  • Quantum communication could allow secure cryptography key distribution and networking of quantum sensors and computers
  • Quantum sensors could make the ocean transparent for Anti-submarine warfare, create unjammable A2/AD, detect stealth aircraft, find hidden tunnels and weapons of mass destruction, etc.
• A few of these technologies are available now, some in the next 5 years and a few are a decade or more out
• Tens of billions of public and private capital dollars are being invested in them
• Defense applications will come first
• The largest commercial applications won’t be the ones we currently think they’re going to be
  • when they do show up they’ll destroy existing businesses and create new ones

Filed under: Gordian Knot Center for National Security Innovation, Technology | 2 Comments »

The last year has seen a ton written about the semiconductor industry: chip shortages, the CHIPS Act, our dependence on Taiwan and TSMC, China, etc.

But despite all this talk about chips and semiconductors, few understand how the industry is structured. I’ve found the best way to understand something complicated is to diagram it out, step by step. So here’s a quick pictorial tutorial on how the industry works.

---

The Semiconductor Ecosystem

We’re seeing the digital transformation of everything. Semiconductors – chips that process digital information — are in almost everything: computers, cars, home appliances, medical equipment, etc. Semiconductor companies will sell $600 billion worth of chips this year.

Looking at the figure below, the industry seems pretty simple. Companies in the semiconductor ecosystem make chips (the triangle on the left) and sell them to companies and government agencies (on the right). Those companies and government agencies then design the chips into systems and devices (e.g. iPhones, PCs, airplanes, cloud computing, etc.), and sell them to consumers, businesses, and governments. The revenue of products that contain chips is worth tens of trillions of dollars.

Yet, given how large it is, the industry remains a mystery to most.  If you do think of the semiconductor industry at all, you may picture workers in bunny suits in a fab clean room (the chip factory) holding a 12” wafer. Yet it is a business that manipulates materials an atom at a time and its factories cost 10s of billions of dollars to build.  (By the way, that wafer has two trillion transistors on it.)

If you were able to look inside the simple triangle representing the semiconductor industry, instead of a single company making chips, you would find an industry with hundreds of companies, all dependent on each other. Taken as a whole it’s pretty overwhelming, so let’s describe one part of the ecosystem at a time.  (Warning –  this is a simplified view of a very complex industry.)

Semiconductor Industry Segments

The semiconductor industry has seven different types of companies. Each of these distinct industry segments feeds its resources up the value chain to the next until finally a chip factory (a “Fab”) has all the designs, equipment, and materials necessary to manufacture a chip. Taken from the bottom up these semiconductor industry segments are:

1. Chip Intellectual Property (IP) Cores
2. Electronic Design Automation (EDA) Tools
3. Specialized Materials
4. Wafer Fab Equipment (WFE)
5. “Fabless” Chip Companies
6. Integrated Device Manufacturers (IDMs)
7. Chip Foundries
8. Outsourced Semiconductor Assembly and Test (OSAT)

The following sections below provide more detail about each of these eight semiconductor industry segments.

Chip Intellectual Property (IP) Cores

• The design of a chip may be owned by a single company, or…
• Some companies license their chip designs – as software building blocks, called IP Cores – for wide use
• There are over 150 companies that sell chip IP Cores
• For example, Apple licenses IP Cores from ARM as a building block of their microprocessors in their iPhones and Computers

Electronic Design Automation (EDA) Tools

• Engineers design chips (adding their own designs on top of any IP cores they’ve bought) using specialized Electronic Design Automation (EDA) software
• The industry is dominated by three U.S. vendors – Cadence, Mentor (now part of Siemens) and Synopsys
• It takes a large engineering team using these EDA tools 2-3 years to design a complex logic chip like a microprocessor used inside a phone, computer or server. (See the figure of the design process below.)

• Today, as logic chips continue to become more complex, all Electronic Design Automation companies are beginning to insert Artificial Intelligence aids to automate and speed up the process

Specialized Materials and Chemicals

So far our chip is still in software. But to turn it into something tangible we’re going to have to physically produce it in a chip factory called a “fab.” The factories that make chips need to buy specialized materials and chemicals:

• Silicon wafers – and to make those they need crystal growing furnaces
• Over 100 Gases are used – bulk gases (oxygen, nitrogen, carbon dioxide, hydrogen, argon, helium), and other exotic/toxic gases (fluorine, nitrogen trifluoride, arsine, phosphine, boron trifluoride, diborane, silane, and the list goes on…)
• Fluids (photoresists, top coats, CMP slurries)
• Photomasks
• Wafer handling equipment, dicing
• RF Generators

Wafer Fab Equipment (WFE) Make the Chips

• These machines physically manufacture the chips
• Five companies dominate the industry – Applied Materials, KLA, LAM, Tokyo Electron and ASML
• These are some of the most complicated (and expensive) machines on Earth. They take a slice of an ingot of silicon and manipulate its atoms on and below its surface
• We’ll explain how these machines are used a bit later on

 “Fabless” Chip Companies

• Systems companies (Apple, Qualcomm, Nvidia, Amazon, Facebook, etc.) that previously used off-the-shelf chips now design their own chips.
• They create chip designs (using IP Cores and their own designs) and send the designs to “foundries” that have “fabs” that manufacture them
• They may use the chips exclusively in their own devices e.g. Apple, Google, Amazon ….
• Or they may sell the chips to everyone e.g. AMD, Nvidia, Qualcomm, Broadcom…
• They do not own Wafer Fab Equipment or use specialized materials or chemicals
• They do use Chip IP and Electronic Design Software to design the chips

Integrated Device Manufacturers (IDMs)

• Integrated Device Manufacturers (IDMs) design, manufacture (in their own fabs), and sell their own chips
  • They do not make chips for other companies (this is changing rapidly – see here.)
  • There are three categories of IDMs– Memory (e.g. Micron, SK Hynix), Logic (e.g. Intel), Analog (TI, Analog Devices)
• They have their own “fabs” but may also use foundries
  • They use Chip IP and Electronic Design Software to design their chips
  • They buy Wafer Fab Equipment and use specialized materials and chemicals
• The average cost of taping out a new leading-edge chip (3nm) is now $500 million

 Chip Foundries

• Foundries make chips for others in their “fabs”
• They buy and integrate equipment from a variety of manufacturers
  • Wafer Fab Equipment and specialized materials and chemicals
• They design unique processes using this equipment to make the chips
• But they don’t design chips
• TSMC in Taiwan is the leader in logic, Samsung is second
• Other fabs specialize in making chips for analog, power, rf, displays, secure military, etc.
• It costs $20 billon to build a new generation chip (3nm) fabrication plant

Fabs

• Fabs are short for fabrication plants – the factory that makes chips
• Integrated Device Manufacturers (IDMs) and Foundries both have fabs. The only difference is whether they make chips for others to use or sell or make them for themselves to sell.
• Think of a Fab as analogous to a book printing plant (see figure below)

1. Just as an author writes a book using a word processor, an engineer designs a chip using electronic design automation tools
2. An author contracts with a publisher who specializes in their genre and then sends the text to a printing plant. An engineer selects a fab appropriate for their type of chip (memory, logic, RF, analog)
3. The printing plant buys paper and ink. A fab buys raw materials; silicon, chemicals, gases
4. The printing plant buys printing machinery, presses, binders, trimmers. The fab buys wafer fab equipment, etchers, deposition, lithography, testers, packaging
5. The printing process for a book uses offset lithography, filming, stripping, blueprints, plate making, binding and trimming. Chips are manufactured in a complicated process manipulating atoms using etchers, deposition, lithography. Think of it as an atomic level offset printing. The wafers are then cut up and the chips are packaged
6. The plant turns out millions of copies of the same book. The plant turns out millions of copies of the same chip

While this sounds simple, it’s not. Chips are probably the most complicated products ever manufactured.  The diagram below is a simplified version of the 1000+ steps it takes to make a chip.

Outsourced Semiconductor Assembly and Test (OSAT)

• Companies that package and test chips made by foundries and IDMs
• OSAT companies take the wafer made by foundries, dice (cut) them up into individual chips, test them and then package them and ship them to the customer

 

Fab Issues

• As chips have become denser (with trillions of transistors on a single wafer) the cost of building fabs have skyrocketed – now >$10 billion for one chip factory
• One reason is that the cost of the equipment needed to make the chips has skyrocketed
  • Just one advanced lithography machine from ASML, a Dutch company, costs $150 million
  • There are ~500+ machines in a fab (not all as expensive as ASML)
  • The fab building is incredibly complex. The clean room where the chips are made is just the tip of the iceberg of a complex set of plumbing feeding gases, power, liquids all at the right time and temperature into the wafer fab equipment
• The multi-billion-dollar cost of staying at the leading edge has meant most companies have dropped out. In 2001 there were 17 companies making the most advanced chips.  Today there are only two – Samsung in Korea and TSMC in Taiwan.
  • Given that China believes Taiwan is a province of China this could be problematic for the West.

What’s Next – Technology

It’s getting much harder to build chips that are denser, faster, and use less power, so what’s next?

• Instead of making a single processor do all the work, logic chip designers have put multiple specialized processors inside of a chip
• Memory chips are now made denser by stacking them 100+ layers high
• As chips are getting more complex to design, which means larger design teams, and longer time to market, Electronic Design Automation companies are embedding artificial intelligence to automate parts of the design process
• Wafer equipment manufacturers are designing new equipment to help fabs make chips with lower power, better performance, optimum area-to-cost, and faster time to market

What’s Next – Business

The business model of Integrated Device Manufacturers (IDMs) like Intel is rapidly changing. In the past there was a huge competitive advantage in being vertically integrated i.e. having your own design tools and fabs. Today, it’s a disadvantage.

• Foundries have economies of scale and standardization. Rather than having to invent it all themselves, they can utilize the entire stack of innovation in the ecosystem. And just focus on manufacturing
• AMD has proven that it’s possible to shift from an IDM to a fabless foundry model. Intel is trying. They are going to use TSMC as a foundry for their own chips as well as set up their own foundry

What’s Next – Geopolitics

Controlling advanced chip manufacturing in the 21st century may well prove to be like controlling the oil supply in the 20th. The country that controls this manufacturing can throttle the military and economic power of others.

• Ensuring a steady supply of chips has become a national priority. (China’s largest import by $’s are semiconductors – larger than oil)
• Today, both the U.S. and China are rapidly trying to decouple their semiconductor ecosystems from each other; China is pouring $100+ billion of government incentives in building Chinese fabs, while simultaneously trying to create indigenous supplies of wafer fab equipment and electronic design automation software
• Over the last few decades the U.S. moved most of its fabs to Asia. Today we are incentivizing bringing fabs and chip production back to the U.S.

An industry that previously was only of interest to technologists is now one of the largest pieces in great power competition.

Filed under: Gordian Knot Center for National Security Innovation, Technology | 22 Comments »