Enterprise 00 01 02 03 04 05 06 07 08 09 10 11 12 Watch Closely Informs Strategy Act Now

Enterprise Trends

The Rise of MLOps Businesses can turn their unruly data- language processing collection and classi- As matures and new sets into structured data that can be fication. Trained to recognize keywords, applied business solutions emerge, devel- trained, and they can build and deploy special algorithms can rapidly sort, opers are shifting their focus from build- models with minimal skills. Create ML classify, and tag information to detect ing models to operating them. Within is Apple’s no-code, drag-and-drop tool patterns. For example, a model trained to software, a set of best practices known as that lets users build custom models such search for hate speech can detect bad ac- DevOps relies on tools, automation, and as recommendation engines, natural tors in social networks. Machine transla- workflows to reduce complexity so that processing systems, and text classifiers. tion generates training data for financial developers can focus on problems that Google’s AutoML includes image clas- crime classification; last year, it reduced need to be solved. This approach is now sification, object detection, translation, the amount of time needed for classifi- being used in machine learning. In 2020, and all sorts of pattern recognition tools. cation from 20 weeks (human analysts some of the fastest-growing GitHub MakeML creates object detection. Ap- working alone) to two weeks. projects were MLOps, or projects that plications have included tracking tennis dealt with tooling, infrastructure, and balls during matches and automatically Simulating Empathy and changing the colors of objects (such as operations. Going forward, MLOps AI can now measure biomarkers that will describe a set of best practices that flowers or dresses) in images. Last year, Amazon launched a no-code mobile suggest a person’s emotional state, combines machine learning, traditional such as agitation, sadness, or giddiness. DevOps, and data engineering. and web app builder for Amazon Web Services (AWS). Microsoft Power Apps Precisely detecting human emotion is is a low-code application development challenging, but companies with a large Low-Code or No-Code Machine environment on Azure. enough dataset are developing accurate By measuring certain biomarkers, AI can detect people’s and respond accordingly. Learning models. Amazon’s Rekognition API infers someone’s emotions using facial Machine learning is transitioning, as new Web-Scale Content Analysis platforms allow businesses to leverage recognition and physical appearance. the power of AI to build applications Mining very large, unstructured datasets Replika uses AI to evaluate voice and without the need to know specific code. is now easier thanks to advanced natural text, and over time it mirrors the user.

12 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now

Enterprise Trends

Affectiva Human Perception AI analyzes theory of mind models of their own. This Serverless Computing complex human states using speech ana- technology could improve existing AI AWS, Alibaba Cloud, Microsoft’s Azure, lytics, , and deep learn- therapy applications such as WoeBot, Google Cloud, and Baidu Cloud are ing. For example, the automotive sector a clinical therapy chatbot. By designing rolling out new offerings and packages uses ’s technology to detect a machines to respond with empathy and for developers with the goal of making driver’s emotional state—such as sleep- concern, digital assistants such as Alexa it easier and more affordable for a wide iness or road rage—and make real-time will increasingly become a part of one’s swath of AI startups to launch their ideas suggestions to improve their driving. family. This technology could eventually into the marketplace. AWS Lambda lets end up in hospitals, schools, and prisons, teams run code for virtually any type of Artificial Emotional Intelligence providing emotional support robots to application or back-end service—with- patients, students, and inmates. Accord- Research teams at Loving AI and Han- out provisioning or managing servers ing to health service organization Cigna, or hands-on administration. The Azure sen Technologies are teaching machines the rate of loneliness in the U.S. has unconditional love, active listening, Functions architecture supports myr- doubled in the past 50 years. Two years iad programming languages, scales on and empathy. In the future, machines ago, former U.K. Prime Minister Theresa will convincingly exhibit human emo- demand, and charges only for active May created a new cabinet position, the compute time. Some engineers worry tions such as love, happiness, fear, and world’s first Minister of Loneliness. In sadness. It begs the question: What is that such serverless systems require them our increasingly connected world, people to surrender too much control. an authentic emotion? Theory of mind report feeling more isolated. Future refers to the ability to imagine the governments struggling with a massive mental state of others. This has long mental health crisis, such as South Korea, Samsung’s next Exynos system on a chip will have an AMD graphics processing unit (GPU). been considered a trait unique to humans may turn to emotional support robots to and certain primates. AI researchers address the issue at scale. are working to train machines to build

13 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12

Expert Insight

Emotion AI Will words we say: We express ourselves through nonverbal cues from our Power the Empathy faces, voices, and body language. But technology is not designed to Economy, but AI capture the nuances of how we Still Needs to Work interact with those around us. AI may be the answer to preserv- ing our humanity in virtual envi- Dr. Rana El Kaliouby ronments. Specifically, Emotion AI—software that can understand CEO of Affectiva nuanced human emotions and complex cognitive states based on facial and vocal expressions— Emotion AI will power the empathy can address some of technology’s economy, but AI still needs to work. shortcomings in light of the pan- demic, and we’ll see companies The COVID-19 pandemic has using it for new use cases, such as: meant that more than ever, we rely on video conferencing to connect 1. Video conferencing virtually–working remotely, learning and virtual events from home, and in our social lives. Emotion AI can provide insight But there’s a big problem: These on how people are emotionally technologies are emotion blind. engaging in a virtual event or meeting. This provides presenters When we communicate in person, with valuable audience feed- we convey so much more than the

© 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

back, gives participants a sense of a patient’s emotional wellbe- What’s on the horizon for AI: advanced driver safety features of shared experience, and can ing provides a quantitative mea- 1. Data synthesis. and personalize transportation, help companies take a pulse on sure of mental health that goes and to achieve that car makers AI algorithms are built on deep collective engagement during beyond self-reporting on a rating want to better understand what’s learning, but they can only work this stressful time. scale of 1-10. happening with people inside of accurately when they’re trained a vehicle. Getting that real-world and then validated on massive 2. Online learning data is difficult, expensive and amounts of data. That includes time-consuming. But data syn- Emotion AI can give feedback data that are diverse and truly thesis is not. For example, a video on how students are engaging representative of the situations of a person driving a car can with online educational ma- the algorithm will encounter in AI researchers are become data that lets research- terials and lectures, flagging if the real world. they’re confused, stressed, or now taking data that’s ers create new scenarios, such as simulating the person turning her bored. This becomes especially But companies developing AI of- already been collected head, or wearing a hat or sun- important during the pandemic ten are challenged in getting ac- glasses. as so many students are learn- and synthesized and cess to the right kinds of data and ing online and suffering from the necessary volumes of data. using it to create brand 2. The need for diversity, equity “Zoom fatigue.” That’s where data simulation and and inclusion. new data. data synthesis methodologies will 3. Telehealth come into play, addressing those As AI becomes more mainstream, Emotion AI can create more problems. AI researchers are now the tech is taking on roles that meaningful discussions and trust taking data that’s already been were traditionally done by hu- between patients and healthcare collected and synthesized and mans and changing how we providers as telehealth appoint- using it to create brand new data. interact with one another. For the ments are replacing in-person technology to work for all of us, visits. And, a data-driven analysis Take the automotive industry. diverse teams must build those The industry seeks to develop applications. Indeed, the number

15 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

one issue to look out for is the This can have adverse implica- ** risk for bias. Unfortunately, we’ve tions for social and economic Dr. Rana El Kaliouby is the Emotion AI can seen many instances in which AI mobility. People with access to has been biased against minority certain types of AI will be able Co-founder and CEO of Affectiva, give feedback on groups. Not only is this unethical; to work more efficiently and the pioneer of Emotion AI. Rana in- how students are it’s also bad for business. If AI will have a leg-up on those who vented the company’s award-win- can’t work for all people as it’s don’t have access. I worry about ning technol- engaging with online intended, there’s little benefit to the impact this can have on ogy. Prior to founding Affectiva, el educational materials using it in the first place. communities and populations Kaliouby was a research scientist that are already disadvantaged, and lectures, flagging at the MIT Media Lab where she 3. The challenge of power because AI could continue to spearheaded applications for if they’re confused, asymmetry. widen that gap. facial coding to benefit mental stressed, or bored. Powerful technologies like AI health, autism, and other research We need to create guidelines to are often in the hands of large areas. Born and raised in Cairo, ensure AI is applied in an equi- corporations and governments, she received degrees in computer table way. The technology has and this poses a number of science from the American Univer- challenges. The value that users the potential to improve people’s sity in Cairo and a Ph.D. from the receive from the technologies lives and solve societal prob- computer laboratory, University of don’t always measure up to the lems, but if we don’t start thinking Cambridge. value that companies gain from about power distribution now, we the tech’s user data. Also, those risk institutionalizing AI in a way corporations or governments can that may exacerbate inequalities. determine a technology’s distri- bution and who has access to it.

16 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now

Enterprise Trends

AI in the Cloud AI at the Edge to train, and have relied on data centers exa’s back-end services rather than chips ations. Siemens MindSphere supports Corporate leaders within the AI eco- AI-driven processing and decision-mak- and computers that consume hundreds designed by Nvidia. The AI chip market digital twins for a number of industries. system have been racing to capture AI ing that occurs closer to the source of data of kilowatts of power. That is all starting will quadruple to $6.7 billion in 2022, As low-code and no-code systems become cloudshare—and to become the most generation, as opposed to in the cloud, is to change. Enter the SoC, or “system on from $1.66 billion in 2018, according to more prevalent, companies should be trusted provider of AI on remote servers. a technique known as “edge computing.” a chip.” Big tech companies including market research firm Tractica. Marketing able to build and deploy digital twins to Enterprise customers are likely to stick The Internet of Things and its billions of Huawei, Apple, Microsoft, Facebook, pretrained chips to businesses will speed simulate a wide array of processes, which with their initial vendor, because machine devices, combined with 5G networking Alphabet, IBM, Nvidia, Intel, and Qual- up commercialization and further R&D. will lead to reduced spending on modern- learning systems get better over time, the and increased computing power, has comm, as well as startups Graphcore, But if the various device manufactur- ization efforts. more data they amass. For that reason, made large-scale AI at the edge possible. Mythic, Wave Computing, SambaNova ers all start creating unique protocols, the competition is furious, even though Processing data directly on devices will be Systems, and Cerebras Systems, are all developers may struggle with too many Spotting Fakes it’s still early. In the West, the field is important in the future for health care, working on new systems architecture and different frameworks. We anticipate an SoCs—some of which come pretrained. In eventual consolidation, pitting just a few In the past year, researchers showed led by Amazon, Microsoft, and Google, automotive, and manufacturing applica- how AI could be used to compose text followed by companies including Apple, tions because it’s potentially faster and short, this means that the chips are more companies—and their SoCs and languag- readily able to work on AI projects and es—against one another. so good that humans couldn’t tell it was IBM, Salesforce, SAP, and Oracle. In safer. Apple spent $200 million to ac- machine written. The team at OpenAI Asian markets, Alibaba and Baidu dom- quire Xnor.ai, a Seattle-based AI startup should promise faster and more secure processing. Projects that might otherwise demonstrated the many reasons why this inate the AI cloud, although in January focused on low-power machine learning Digital Twins was problematic, from mass-generating 2020, telecom equipment and smartphone software and hardware. Microsoft offers take weeks could instead be accomplished in a matter of hours. Cerebras has built Digital twins are virtual representations salacious social media posts and fake maker Huawei announced a management a comprehensive toolkit called Azure of real-world environments, products, reviews to forging documents by world change to focus on what it calls a “full- IoT Edge that allows AI workloads to be an AI chip with 1.2 trillion transistors, 400,000 processor cores, 18 gigabytes of or assets for a variety of purposes. Man- leaders. It turns out that AI can also be stack cloud platform.” It’s a $250 billion moved to the edge. ufacturers use digital twins to manage used to detect when text was machine industry and quickly growing. New York SRAM, and interconnects (tiny connec- tion nodes) that can move 100 quadrillion the performance and effectiveness of generated, even if we humans can’t spot University Stern School of Business Advanced AI Chipsets machines and plants, while city planners the fake. That’s because an essay written professor Arun Sundararajan says it best: bits per second. That’s an astounding Today’s neural networks have long amount of components and power. As of use them to simulate the impact of new by AI tends to rely on statistical patterns “The prize will be to become the operat- developments and roads. The Singapore in text and doesn’t have much linguistic ing system of the next era of tech.” required an enormous amount of com- November 2020, Amazon’s homegrown puting power, have taken a long time AI chip AWS Inferentia now powers Al- government uses them for urban oper- variation. Researchers at the MIT-IBM

17 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now

Enterprise Trends

Watson AI Lab and Harvard University Intelligent Optical Character among enterprise companies. Google’s developed the Giant Language Mod- Recognition Duplex is a good example; it’s a bot de- el Test Room (GLTR), which looks An ongoing challenge is getting machines signed to make routine phone calls. Ama- for words that are likely to appear in a to recognize the various ways we express zon uses RPA to sift through résumés and particular order. This technology can be ourselves in writing. Optical character prioritize top candidates. In banking, Blue used to detect forgery, intentional records recognition (OCR) works in fixed, recog- Prism and Automation Anywhere help falsification, email phishing campaigns, nizable formats such as highway signs staff with repetitive work functions. RPA and corporate espionage. and the text from a book. But often, OCR will eventually augment staff and shift isn’t smart enough to recognize different productivity into higher gear. Natural Language Processing fonts, unique notations, or spreadsheets for ESGs with fields specific only to one company. Massive Translation Systems Companies are moving toward new envi- Researchers are training AI systems to In 2020, Facebook launched a new ronmental, social, and governance (ESG) recognize patterns, even if they show up open-source AI language model called criteria—a set of standards increasingly in unusual places. For example, the AWS M2M-100 that can translate 100 languag- used by investors to evaluate their invest- Textract system now recognizes both es. Facebook’s AI lab trained the model ments. ESG standards must be quanti- text and context specific to a company or using 7.5 billion sentence pairs gathered fied and explicitly stated, but measuring business unit. automatically from the web. (Surpris- performance can be difficult because ingly, Facebook did not use its own data many intangibles or abstract concepts are Robotic Process Automation for this project.) The FastText language involved. Natural language processing is Robotic process automation (RPA) can model identified the language, and an Facebook launched the first AI model that translates 100 languages without relying on English data. being used to identify, tag, and sort doc- unsupervised learning model matched (Image credit: Facebook.) automate certain tasks and processes umentation from various sources about a within offices and allow employees to sentences by their meaning. The goal company’s ESG reputation (on issues such spend time on higher-value work. It’s the was to improve simultaneous language as labor practices, community impact, most commonly deployed AI technique translation. diversity, and inclusion).

18 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now

Enterprise Trends

Predicting Systems and misinterprets data and neglects to identify is news, products, or advertising. This their work. In 2018, Microsoft paid $7.5 100-Year Software Site Failures cancer among certain patients? These are resulted in the ongoing antitrust lawsuits billion to buy GitHub, a popular devel- Traditional software has a short and Computer vision can anticipate and iden- the kinds of problems that could put a filed against the companies. opment platform allowing anyone to unpredictable shelf life compared with tify failures in physical locations. High- company at risk of lawsuits. New insur- host and review code, to collaborate with other engineering tools. This leads to tech factories, airline manufacturers, and ance models will help address these issues. Global Rush to Fund AI other developers, and to build all kinds headaches and costly upgrades, often with construction sites use image recognition Underwriters are starting to include AI of projects. AWS hosts its own market- downtime. As a result, companies and under cyber insurance plans. Specialty There is a global race to fund AI research place, offering models and algorithms for systems to monitor projects and automat- and to acquire AI startups. In the first government agencies attempt to keep ically warn of problems. This is accom- insurers such as LaPlaya Insurance now computer vision, speech recognition, and pace with the evolution of technology by offer insurance for AI applications. quarter of 2020, 285 U.S.-based AI start- text—and its base of sellers includes Intel, plished by comparing data from the real ups had raised $6.9 billion, according to maintaining systems rather than evolv- world to that of a digital twin. CloudSight, and many others. (Think ing. Libraries, data formats, and protocols the National Venture Capital Association. of AWS Marketplace as an Amazon for Manipulating AI Systems for Investment waned as Covid became a can all become outdated quickly, creating Competitive Advantage algorithms and models.) There are mar- vulnerabilities in critical systems. Since Liability Insurance for AI global pandemic, but tech giants includ- ketplaces for generalists, like GenesisAI Amazon, Google, and Facebook have ing Apple, Google, and Microsoft are still 2015, the Defense Advanced Research Who’s to blame when machines behave and Algorithmia, where developers can Projects Agency (DARPA) has funded badly? When the machine learning all come under fire in the past few years acquiring AI companies, while non-tech upload their work and receive payment for manipulating their search systems to companies are gobbling AI startups too: research to make software viable for system in Uber’s self-driving car failed when others pay to access it. Now there more than 100 years. These systems and killed an Arizona pedestrian, the prioritize results that are more profitable McDonald’s acquired personalization are specialized marketplaces for specific for their companies. For example, Google platform Dynamic Yield, while Nike would use AI to dynamically adapt to company was likely not covered under use cases: Nuance AI Marketplace devel- changes in environments and resources. traditional cyber insurance. As busi- has been accused of de-ranking websites acquired inventory management compa- oped a single API to connect its algo- and promoting news stories from pre- ny Celect and guided shopping experience They require a novel approach to design, nesses rush to build and implement AI rithms to radiologists at 6,500 health care using AI to discover and make visible the products and processes, they must plan ferred partners. Late in 2019, researchers platform Invertex. facilities. Quantiacs allows developers to found that Amazon had optimized its application’s operations and interactions ahead for emerging risks. For example, build algorithmic trading systems, and with other systems. what happens if machine learning makes search algorithm to boost the visibility of Algorithm Marketplaces it matches their algorithms with capital Amazon’s own brands. Tweaks to search a company vulnerable to attackers who In the 2010s, big tech companies, start- from institutional investors. Bonseyes is a inject fake training data into a system? algorithms have a significant impact on European-specific marketplace to buy and what internet users see, whether that ups, and communities of developers used What if a health care company’s AI algorithm marketplaces to share and sell sell AI tools.

19 © 2021 Future Today Institute 00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence

Rage Against the Machine

Mid-future neutral scenario From screaming into pillows to pounding punching bags, humans have developed numerous ways to air our frustrations. The hope is that if we act out against inani- mate objects, we’re less likely to act out against our fellow humans, risking harm or trauma. But what if there were a humanoid stand-in that could absorb our aggressions in a more cathartic, and ultimately beneficial way? As AI begins to achieve convincing emulations of human per- sonalities, a new type of avatar emerges, algorithmically designed to provide a responsive therapeutic outlet for aggression. Users can program the avatar to look and act like a figure from their life (a boss, a partner, a rival) for whom they harbor pent-up feelings, allowing users to express them- selves freely without threatening their real-life relation- ships or risking legal repercussions. The AI persona could even be assigned to a surrogate robotic body, letting the user act out physical aggressions. But as the technology grows in popularity, designers must keep watch that what they’ve created doesn’t normalize and increase the rate of interhuman conflict, instead of alleviating it.

Scenarios20 © 2021 Future Today Institute