2021 Tech Trends Report Is De- It, Produced Two Factions: Those Signed to Help You Confront Deep Who Wanted to Reverse Time and Uncertainty, Adapt and Thrive

Volume 1 of 12

14th Annual Edition 2021 Tech Trends

Report Artificial Strategic trends that will influence business, Intelligence government, education, media and society in the coming year. 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

03 Overview 19 Algorithm Marketplaces 32 Machine Image Completion 43 Nation-based Guardrails 04 Macro Forces and Emerging Trends 19 100-Year Software 32 Predictive Models Using and Regulations 06 Summary 20 Scenario: Rage Against the Machine Single Images 43 Regulating Deepfakes 08 Artificial Intelligence 22 Health, Medicine, and Science 33 Model-free Approaches to RL 43 Making AI Explain Itself 09 An Executive’s Guide to AI 22 AI Speeds Scientific Discovery 33 Real-time Machine Learning 43 New Strategic Technical Alliances 09 Machine Learning 22 AI-First Drug Discovery 33 Automated Machine Learning 43 The New Mil-Tech (AutoML) Industrial Complex 09 Deep Learning 23 AI Improves Patient Outcomes 33 Hybrid Human-Computer Vision 44 Algorithmic Warfighting 10 Weak and Strong AI 23 Deep Learning Applied to 33 Neuro-Symbolic AI 12 Enterprise Medical Imaging 33 General Reinforcement 12 The Rise of MLOps 23 NLP Algorithms Detect Virus Mutations Learning Algorithms 12 Low-Code or No-Code 23 Diagnostics Without Tests 34 Continuous Learning Machine Learning 45 China’s AI Rules 34 Proliferation of 12 Web-Scale Content Analysis 23 Protein Folding Franken-Algorithms 48 Society 12 Simulating Empathy and Emotion 23 Dream Communication 34 Proprietary, Homegrown 48 Ethics Clash 24 Thought Detection 13 Artificial Emotional Intelligence AI Languages 48 Ambient Surveillance 25 Scenario: Deep Twins in the OR 13 Serverless Computing 36 Talent 48 Marketplace Consolidation 27 Consumer 14 Expert Insight: Emotion AI Will 36 AI Brain Drain 48 Fragmentation Power the Empathy Economy, 27 Zero UIs 36 AI Universities 49 Expert Insight: AI reveals but AI Still Needs to Work 27 Consumer-grade AI Applications 37 Demand for AI Talent Growing our real-world biases 17 AI in the Cloud 27 Ubiquitous Digital Assistants 37 Corporate AI Labs 50 AI Still Has a Bias Problem 17 AI at the Edge 28 Deepfakes for Fun 37 AI for Interviews 50 Problematic Training Data 17 Advanced AI Chipsets 28 Personal Digital Twins 39 Creative 50 AI to Catch Cheaters 17 Digital Twins 30 Research 39 Assisted Creativity 50 Algorithms Targeting 17 Spotting Fakes 30 Closed-Source Code Vulnerable Populations 39 Generative Algorithms for 18 Natural Language Processing 30 Framework Consolidation Content Production 51 AI Intentionally Hiding Data for ESGs 30 Cost of Training Models 39 Generating Virtual Environments 51 Undocumented AI Accidents 18 Intelligent Optical Character 31 NLP Benchmarks from Short Videos 51 Digital Dividends Recognition 31 Machine Reading Comprehension 40 Automated Versioning 51 Prioritizing Trust 18 Robotic Process Automation 31 AI Summarizing Itself 40 Automatic Voice Cloning 52 Scenario: Bully Bots 18 Massive Translation Systems 31 No Retraining Required and Dubbing 53 Application 19 Predicting Systems and 40 Automatic Ambient Noise Dubbing Site Failures 31 Graph Neural Networks 54 Key Questions 42 Geopolitics and Defense 19 Liability Insurance for AI 31 Federated Learning 55 Sources 42 AI Nationalism 19 Manipulating AI Systems for 31 GP Models 56 Authors Competitive Advantage 31 GPT-3’s Influence 42 National AI Strategies 19 Global Rush to Fund AI 32 Vokenization 42 AI as Critical Infrastructure 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

Overview

The 1920s began in chaos. Cata- It’s difficult not to see striking ed trends. In total, we’ve analyzed clysmic disruption resulting from parallels to our modern world. A nearly 500 technology and science the first world war and the Spanish tumultuous U.S. election, extreme trends across multiple industry flu shuttered businesses and pro- weather events and Covid-19 sectors. In each volume, we discuss voked xenophobia. Technological continue to test our resolve and the disruptive forces, opportunities marvels like the radio, refrigerator, our resilience. Exponential tech- and strategies that will drive your vacuum cleaner, moving assembly nologies—artificial intelligence, organization in the near future. line and electronic power trans- synthetic biology, exascale com- Now, more than ever, your organi- mission generated new growth, puting, autonomous robots, and zation should examine the poten- even as the wealth gap widened. off-planet missions to space—are tial near and long-term impact of More than two-thirds of Ameri- challenging our assumptions about tech trends. You must factor the cans survived on wages too low to human potential. Under lockdown, trends in this report into your stra- sustain everyday living. The pace we’ve learned how to work from tegic thinking for the coming year, of scientific innovation—the dis- our kitchen tables, lead from our and adjust your planning, opera- covery of insulin, the first modern spare rooms, and support each tions and business models accord- antibiotics, and insights into theo- other from afar. But this disruption ingly. But we hope you will make retical physics and the structure of has only just begun. time for creative exploration. From atoms—forced people to reconsid- With the benefit of both hindsight chaos, a new world will come. er their cherished beliefs. and strategic foresight, we can The sheer scale of change, and the choose a path of reinvention. Our great uncertainty that came with 2021 Tech Trends Report is de- it, produced two factions: those signed to help you confront deep who wanted to reverse time and uncertainty, adapt and thrive. For Amy Webb return the world to normal, and this year’s edition, the magnitude Founder those who embraced the chaos, of new signals required us to cre- The Future Today Institute faced forward, and got busy build- ate 12 separate volumes, and each ing the future. report focuses on a cluster of relat-

1 Macro Forces and

Emerging Trends 2 3 4 5

For nearly two decades, the Future Today Institute has meticulously re- searched macro forces of change and the emerging trends that result. Our focus: understanding how these forces and trends will shape our futures. Our 14th annual Tech Trends Report identifies new opportunities for growth and potential collaborations in and adjacent to your business. We also highlight emerging or atypical threats across most industries, including all levels of government. For those in creative fields, you will find a wealth of new ideas that will spark your imagination.

Our framework organizes nearly 500 trends into 12 clear categories.

Within those categories are specific use cases and recommendations for key roles in many organizations: strategy, innovation, R&D, and risk.

Each trend offers six important insights.

1. Years on the List 2. Key Insight 4. Disruptive Impact 6. Action Scale Informs Strategy We track longitudinal tech and Concise description of this trend The implications of this trend on FTI’s analysis of what action your Strong evidence and data. Longer- science trends. This measurement that can be easily understood and your business, government, or organization should take. Fields term uncertainties remain. Use it to indicates how long we have repeated to others. society. include: inform your strategic planning. followed the trend and its progression. 3. Examples 5. Emerging Players Watch Closely Act Now Real-world use cases, some of Individuals, research teams, Mounting evidence and data, but Ample evidence and data. This which should be familiar to you. startups, and other organizations more maturity is needed. Use it to trend is already mature and emerging in this space. inform your vision, planning, and requires action. research. 04 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

Macro Forces and Emerging Trends

Scenarios Describe Plausible Outcomes

You will find scenarios imagining future worlds as trends evolve and converge. Scenarios offer a fresh perspective on trends and often challenge your deeply held beliefs. They prompt you to consider high-impact, high-uncertainty situations using signals available today. 1

1. Headline 2 A short description offering you a glimpse into future changes.

2. Temporal and Emotive Tags 3 A label explaining both when in the future this scenario is set and whether it is optimistic, neutral, pessimistic, or catastrophic.

3. Narrative The descriptive elements of our imagined world, including the developments leading us to this point in our future history.

Scenario sources: The Future Today Institute uses a wide array of quali- tative and quantitative data to create our scenarios. Some of our typical sources include patent filings, academic preprint servers, archival research, policy briefings, conference papers, data sets, structured interviews with experts, conversations with kids, critical design, and specula- tive fiction.

+ Natural language processing is + Natural language processing an area experiencing high inter- algorithms— typically used for est, investment, and growth. text, words, and sentences—are being used to interpret genetic + No-code or low-code systems changes in viruses. are unlocking new use cases for businesses. + COVID-19 accelerated the use of AI in drug discovery last year. The + Amazon Web Services, Azure, first trial of an AI-discovered drug and Google Cloud’s low-code is underway in Japan. and no-code offerings will trickle down to everyday people, al- + AI plays key roles in synthetic Artificial lowing them to create their own biology, genetics, and medical artificial intelligence applica- imaging; predicting the spread of tions and deploy them as easily disease; and improving patient as they could a website. health outcomes.

+ The race is on to capture AI + New artificial nervous systems cloudshare—and to become the use AI and neural implants. Intelligence most trusted provider of AI on remote servers. + The SuperGLUE benchmark, which measures AI’s human + The AI community still operates language ability, will likely be using a closed-source model. Re- surpassed by the end of 2021. searchers’ reluctance to publish Summary their full code leads to less trans- parency and reproducibility, and makes accountability murky. 06 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

+ Gaussian processes, the gold + Technical alliances that help standard for many real-world drive future R&D could also modeling problems, are becom- challenge existing geopolitical ing more accurate and easier alliances. to train. + Future wars will be fought in + AI researchers are leaving ac- code, using data and algorithms ademia for corporations at an as powerful weapons. alarming pace. + We continue to fail to see + Generative adversarial networks China’s growing AI proficiency assist artists and musicians in as a military, economic, and Artificial new forms of creative expression. diplomatic threat. + A new wave of AI nationalism is + New software could be viable for rising as governments institute 100 years by using AI to adapt to new restrictions on M&A and changes around it. investment activity.

Intelligence + Several countries will launch national AI strategies in 2021 and 2022.

+ New measures to regulate the creation and distribution of Summary deepfakes will be introduced throughout 2021 and 2022.

14TH YEAR ON THE LIST Artificial Intelligence

KEY INSIGHT EXAMPLES DISRUPTIVE IMPACT EMERGING PLAYERS AI is now used across most industries. The convergence of groundbreaking • Broad Institute Artificial intelligence It solves business problems, detects research, business use cases, the explosive • Clarifai represents the third era fraud, improves crop yields, manages growth of data, and improvements in • Clearview AI supply chains, recommends products, computing power and storage are en- of computing, generally • DeepMind and even assists designers and writers in abling advances in AI. The global artificial defined as the ability their work. AI can predict call volume in intelligence market is expected to grow at • Disperse for a machine to per- customer service centers and recommend a compound annual growth rate of 42.2% • Graphcore form cognitive functions staffing levels; it also predicts the emo- from 2021 to 2027. • HiSilicon Technologies tional state and behavior of the person • Kasisto calling to help companies anticipate desir- as well as or better than • LabGenius able solutions. AI automates the process humans. Such functions • Mohamed bin Zayed University for drug discovery, which ultimately led of Artificial Intelligence include perception, to faster COVID-19 vaccine candidates. learning, reasoning, Because AI is so broad, we have identified • Niantic problem-solving, con- different themes within the discipline • Nvidia that you should be following. You will • OpenAI textual understanding, also find the technology intersecting with • OpenMined other trends throughout this report. making inferences and • Persado predictions, and exer- • PolyAI cising creativity. • Recursion AI is a force multiplier for every industry. • SenseTime • Scale AI • Syntiant

An Executive’s Guide to AI

What You Need To Know Machine Learning teams know how to classify the input ing to learn about an environment (such themselves. For example, once a system data and what they are trying to predict as a complex financial portfolio), or when learns what an object looks like—say, an In its most basic form, artificial intelli- AI pioneer Arthur Samuel popularized but can get accurate results much more the researcher needs to find greater levels apple—it can recognize that object in all gence is a system that makes autonomous the idea of machine learning in 1959, quickly by relying on an algorithm rather of optimization. It has a tremendous other images, even if it has only a partial decisions. AI is a branch of computer explaining how computers could learn than a human. Understanding what number of business use cases, ranging view. science in which computers are pro- without being explicitly programmed. product features would most likely drive from real-time dynamic pricing models grammed to do things that normally This would mean developing an al- There are different types of deep learning new purchases is an example of a busi- to high frequency trading algorithms to require human intelligence. This includes gorithm that could someday extract models. The most common types include ness use case for supervised learning. the systems that run self-driving cars. learning, reasoning, problem-solving, patterns from datasets and use those convolutional neural networks, recurrent understanding language, and perceiving patterns to predict and make real-time In unsupervised learning, data is pro- neural networks, transformer neural a situation or environment. AI is an decisions automatically. It took many vided to an algorithm without specific Deep Learning networks, and generative adversarial extremely large, broad field that uses years for reality to catch up with Samu- output parameters. For example, if a Deep learning is a relatively new branch networks (GANs). its own computer languages and relies el’s idea, but today machine learning is a researcher doesn’t know quite what to of machine learning. Programmers use A convolutional neural network on computer networks modeled on our primary driver of the growth in AI. do with a large dataset, an algorithm special deep learning algorithms along- (CNN) is multilayered, with a convolu- human brains. Machine learning uses data to make pre- could determine patterns, classify data, side an enormous corpus of data—typ- tional layer, a pooling layer, and a fully dictions and recommendations on how and make recommendations without a ically many terabytes of text, images, connected layer. Each one performs a to achieve stated goals. Types of machine human supervisor. Unsupervised learn- videos, speech, and the like. Often, these different task using the data. The output learning include supervised, unsuper- ing has been used during the pandemic to systems are trained to learn on their is classification. If a researcher has 10,000 vised, and reinforcement. find patterns in how the virus is spread- own, and they can sort through a variety images and needs to extract data—to rec- ing throughout communities. In supervised learning, an algorithm of unstructured data, whether it’s making ognize particular faces, for instance—the In reinforcement learning uses training data to learn the relation- , an algo- sense of typed text in documents or audio CNN would run until information could ship between established parameters— rithm learns to perform a task by repeat- clips or video. In practical terms, deep be inferred. In business, CNNs are used inputs and outputs. Humans supervise, edly running calculations as it attempts to learning’s emergence means that more for recognition: anomalies in medical tweaking and adjusting systems as they accomplish a stated goal. Reinforcement and more human processes will be auto- imaging, faulty products on a production work. Supervised learning is used when learning is used when there isn’t enough mated, including the writing of software, line, blight on crops. training data, when the researcher is try- which computers will soon start to do

An Executive’s Guide to AI

Recurrent neural networks (RNNs) in the past year, are generated using Weak and Strong AI AI. This is problematic for researchers inforcement learning, in which machines are multilayered neural networks that GANs. In design, GANs are tremendous- There are two kinds of AI—weak (or covering AI developments and for man- learn not unlike we do—by trial and move and store information between ly useful: They can produce thousands “narrow”) and strong (or “general”). Nar- agers who must make decisions about error. While we haven’t seen an anthro- input, hidden, and output layers. They of designs and recommend the best ones row AI systems make decisions within AI. In fact, we have already started to pomorphic AI walk out of DeepMind’s are good at modeling sequence data for based on desired parameters. They can very narrow parameters at the same level see real-world examples of functioning lab, we should consider these projects predictions. In business, they are used generate and modulate voices, faces, as a human or better, and we use them all artificial general intelligence. In 2017, as part of a long transition between the anytime the sequence of data matters, even gestures. Researchers from Nvidia, day long without even realizing it. The researchers at DeepMind, a lab owned narrow AI of today and the strong AI of such as speech recognition and language Massachusetts General Hospital, BWH anti-lock brakes in your car, the spam by the same parent company as Google, tomorrow. announced that AI had taught itself how translation. RNNs are used in digital Center for Clinical Data Science, and filter and autocomplete functions in assistants, to create captions for images, the Mayo Clinic collaborated on a GAN your email, and the fraud detection that to play chess, shogi (a Japanese version of and to generate narrative reports using that generates synthetic MRIs showing authenticates you for a credit card pur- chess), and Go (an abstract strategy board structured data (sports, financial). cancerous tumors. chase—these are all examples of artificial game)—all without any human inter- GANs are unsupervised deep learning A transformer is a type of neural net- narrow intelligence. vention. The system, named AlphaZero, systems composed of two competing work architecture that learns what words quickly became the strongest player in Artificial general intelligence (AGI) de- history for each game. The team has neural networks—a generator and a mean when they appear in a particular scribes systems capable of decision-mak- discriminator—that are trained on the context. Using “attention mechanism,” a been publishing important discoveries at ing outside of narrow specialties. Dolores an impressively fast pace. Last year, the same data, such as images of people. The transformer looks at an input sequence in “Westworld,” the Samantha operating networks compete against each other to and determines at each step what other DeepMind team taught AI agents to play system in “Her,” and the H.A.L. super- complex games, such as the capture the perform a task—identifying the correct parts of the sequence are important. To computer from “2001: A Space Odyssey” person—which results in optimizing date, transformers have mainly been flag “game mode” inside the video game are anthropomorphized representa- “Quake III Arena.” They, like humans, overall performance. GANs are useful used in natural language processing and tions of AGI—but the actual technology when researchers don’t have enough data generation. had learned skills specific to the game doesn’t necessarily require humanlike as well as when and how to collaborate to train an algorithmic model. They are appearances or voices. also used to create new, synthetic data. with other teammates. The AI agents had Deepfakes, which have become popular There is no single standard that marks matched human player ability using re- the distinction between weak and strong