Volume 1-Tech - Trends - Volume - 1 - AI - Shared - by - VN - Market - Report - Community

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
Medical Imaging 33 Neuro-Symbolic AI
12 Enterprise
23 NLP Algorithms Detect Virus 33 General Reinforcement
12 The Rise of MLOps
Mutations Learning Algorithms
12 Low-Code or No-Code
23 Diagnostics Without Tests 34 Continuous Learning
Machine Learning 45 China’s AI Rules
23 Protein Folding 34 Proliferation of
12 Web-Scale Content Analysis 48 Society
Franken-Algorithms
12 Simulating Empathy and Emotion 23 Dream Communication
34 Proprietary, Homegrown 48 Ethics Clash
13 Artificial Emotional Intelligence 24 Thought Detection
AI Languages 48 Ambient Surveillance
13 Serverless Computing 25 Scenario: Deep Twins in the OR
36 Talent 48 Marketplace Consolidation
14 Expert Insight: Emotion AI Will 27 Consumer
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 51 AI Intentionally Hiding Data
Content Production
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
31 Graph Neural Networks 40 Automatic Ambient Noise Dubbing 54 Key Questions
Site Failures
31 Federated Learning 42 Geopolitics and Defense 55 Sources
19 Liability Insurance for AI
31 GP Models 42 AI Nationalism 56 Authors
19 Manipulating AI Systems for
Competitive Advantage 31 GPT-3’s Influence 42 National AI Strategies
19 Global Rush to Fund AI 32 Vokenization 42 AI as Critical Infrastructure
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-
03 © 2021 Future Today Institute

00 01 02 03 04 05 06 07 08 09 10 11 12
6
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.
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 chal-
lenge 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 re-
search, policy briefings, conference papers, data sets, structured inter-
views 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
Artificial
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
lowing them to create their own biology, genetics, and medical
Intelligence
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.
most trusted provider of AI on
Summary
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
their full code leads to less trans-
parency and reproducibility, and
makes accountability murky.
+ 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
Artificial
+ 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
new forms of creative expression. diplomatic threat.
Intelligence
+ 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.
+ Several countries will launch
Summary
national AI strategies in 2021
and 2022.
+ New measures to regulate the

creation and distribution of
deepfakes will be introduced
throughout 2021 and 2022.

00 01 02 03 04 05 06 07 08 09 10 11 12 Artificial Intelligence Watch Closely Informs Strategy Act Now
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
staffing levels; it also predicts the emo- from 2021 to 2027. • HiSilicon Technologies
form cognitive functions tional state and behavior of the person • Kasisto
as well as or better than calling to help companies anticipate desir-
• LabGenius
humans. Such functions able solutions. AI automates the process
• Mohamed bin Zayed University
for drug discovery, which ultimately led
include perception, to faster COVID-19 vaccine candidates.
of Artificial Intelligence
• Niantic
learning, reasoning, Because AI is so broad, we have identified
different themes within the discipline • Nvidia
problem-solving, con- that you should be following. You will • OpenAI
textual understanding, also find the technology intersecting with • OpenMined
making inferences and other trends throughout this report.
• 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
AI pioneer Arthur Samuel popularized 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-
the idea of machine learning in 1959, 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
explaining how computers could learn 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
without being explicitly programmed. than a human. Understanding what number of business use cases, ranging view.
science in which computers are pro-
This would mean developing an al- product features would most likely drive from real-time dynamic pricing models There are different types of deep learning
grammed to do things that normally
gorithm that could someday extract new purchases is an example of a busi- to high frequency trading algorithms to models. The most common types include
require human intelligence. This includes
patterns from datasets and use those ness use case for supervised learning. the systems that run self-driving cars. convolutional neural networks, recurrent
learning, reasoning, problem-solving,
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 networks (GANs).
Deep learning is a relatively new branch
its own computer languages and relies el’s idea, but today machine learning is a researcher doesn’t know quite what to A convolutional neural network
of machine learning. Programmers use
on computer networks modeled on our primary driver of the growth in AI. do with a large dataset, an algorithm (CNN) is multilayered, with a convolu-
special deep learning algorithms along-
human brains. Machine learning uses data to make pre- could determine patterns, classify data, tional layer, a pooling layer, and a fully
side an enormous corpus of data—typ-
dictions and recommendations on how and make recommendations without a connected layer. Each one performs a
ically many terabytes of text, images,
to achieve stated goals. Types of machine human supervisor. Unsupervised learn- different task using the data. The output
videos, speech, and the like. Often, these
learning include supervised, unsuper- ing has been used during the pandemic to is classification. If a researcher has 10,000
systems are trained to learn on their
vised, and reinforcement. find patterns in how the virus is spread- images and needs to extract data—to rec-
own, and they can sort through a variety
ing throughout communities. ognize particular faces, for instance—the
In supervised learning, an algorithm of unstructured data, whether it’s making
uses training data to learn the relation- In reinforcement learning, 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.
translation. RNNs are used in digital Center for Clinical Data Science, and filter and autocomplete functions in announced that AI had taught itself how
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

Enterprise
Enterprise Trends
The Rise of MLOps Businesses can turn their unruly data- language processing collection and classi-
As machine learning 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 Emotion
operations. Going forward, MLOps changing the colors of objects (such as
flowers or dresses) in images. Last year, AI can now measure biomarkers that
will describe a set of best practices that suggest a person’s emotional state,
combines machine learning, traditional Amazon launched a no-code mobile
and web app builder for Amazon Web such as agitation, sadness, or giddiness.
DevOps, and data engineering. Precisely detecting human emotion is
Services (AWS). Microsoft Power Apps
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 emotions and respond accordingly.
Learning models. Amazon’s Rekognition API
Machine learning is transitioning, as new infers someone’s emotions using facial
Web-Scale Content Analysis recognition and physical appearance.
platforms allow businesses to leverage
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.

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, computer vision, 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 Affectiva’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- out provisioning or managing servers
Research teams at Loving AI and Han- ing to health service organization Cigna,
sen Technologies are teaching machines or hands-on administration. The Azure
the rate of loneliness in the U.S. has Functions architecture supports myr-
unconditional love, active listening, doubled in the past 50 years. Two years
and empathy. In the future, machines iad programming languages, scales on
ago, former U.K. Prime Minister Theresa demand, and charges only for active
will convincingly exhibit human emo- May created a new cabinet position, the
tions such as love, happiness, fear, and compute time. Some engineers worry
world’s first Minister of Loneliness. In that such serverless systems require them
sadness. It begs the question: What is our increasingly connected world, people
an authentic emotion? Theory of mind to surrender too much control.
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

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

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
Emotion AI can give feedback time-consuming. But data syn-
data that are diverse and truly
on how students are engaging thesis is not. For example, a video
representative of the situations
with online educational ma- of a person driving a car can
the algorithm will encounter in
terials and lectures, flagging if AI researchers are the real world.
become data that lets research-
ers create new scenarios, such as
they’re confused, stressed, or now taking data that’s simulating the person turning her
bored. This becomes especially But companies developing AI of-
important during the pandemic already been collected head, or wearing a hat or sun-
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
“Zoom fatigue.”
using it to create brand the necessary volumes of data.
2. The need for diversity, equity
That’s where data simulation and
new data. data synthesis methodologies will
and inclusion.
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

one issue to look out for is the This can have adverse implica- **
risk for bias. Unfortunately, we’ve tions for social and economic
seen many instances in which AI mobility. People with access to Dr. Rana El Kaliouby is the
Emotion AI can 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-
vented the company’s award-win-
how students are it’s also bad for business. If AI will have a leg-up on those who
can’t work for all people as it’s don’t have access. I worry about ning emotion recognition 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, at the MIT Media Lab where she
and lectures, flagging 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
We need to create guidelines to
health, autism, and other research
are often in the hands of large areas. Born and raised in Cairo,
corporations and governments, ensure AI is applied in an equi-
she received degrees in computer
and this poses a number of table way. The technology has
science from the American Univer-
challenges. The value that users the potential to improve people’s
lives and solve societal prob- sity in Cairo and a Ph.D. from the
receive from the technologies
lems, but if we don’t start thinking computer laboratory, University of
don’t always measure up to the
about power distribution now, we Cambridge.
value that companies gain from
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.

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

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- automate certain tasks and processes unsupervised learning model matched
(Image credit: Facebook.)
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).

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
a company vulnerable to attackers who Amazon’s own brands. Tweaks to search from institutional investors. Bonseyes is a
algorithms have a significant impact on In the 2010s, big tech companies, start-
inject fake training data into a system? ups, and communities of developers used European-specific marketplace to buy and
What if a health care company’s AI what internet users see, whether that sell AI tools.
algorithm marketplaces to share and sell

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.

Health, Medicine
& Science
Health, Medicine, and Science Trends
AI Speeds Scientific Discovery AI-First Drug Discovery maceutical company has inked deals with
Running experiments with several COVID-19 accelerated the use of AI in AI drug discovery startups, including
variables often requires tiny, methodical drug discovery. An international team Johnson & Johnson, Novartis, Merck,
tweaks to measurements, materials, and crowdsourced a Covid antiviral by syn- AstraZeneca, and GlaxoSmithKline.
inputs. Graduate students might spend thesizing candidates for 2,000 molecules Much of the potential in AI stems from
hundreds of tedious hours repeatedly in less than 48 hours—a process that like- deep learning’s ability to sort through
making small adjustments until they find ly would have taken human researchers a huge volumes of information and learn
a solution—a waste of their cognitive abil- month or longer. In Japan, the first phase and extrapolate from that information.
ities. Research labs now use AI systems to of a clinical trial for an AI-designed drug The upshot: AI can think faster than
speed the process of scientific discovery. to treat obsessive-compulsive disorder humans—sorting data in months versus
Materials scientists at the University of showed a positive result. The drug, DSP- years—and see patterns that we may not.
British Columbia now rapidly test a new 1181, acts as an agonist to the receptor Still, drug discovery is tricky, because
kind of solar cell and log results using a for serotonin, a signaling molecule in the the algorithms rely on drug targets that
robot overseen by an AI algorithm. Based brain that mediates mood. The project must be published in research journals.
on the results of each experiment, an al- used AI techniques to generate tens of Most data about potential compounds
gorithm determines what to change next. millions of potential molecules to try isn’t readily available, and when it is, it
A 9- to 12-month process was completed against the serotonin receptor and sift isn’t always complete or reliable. Filling
in five days. Google’s DeepMind devel- through the candidates to decide which the gaps and cleaning that data takes time
oped a way of testing and modeling the ones to prioritize for synthesis and and money. It also requires data shar-
complex folding patterns of long chains testing. ing—and most drug data is proprietary
The U.S. Food and Drug Administration approved IDx-DR, the first autonomous AI system to provide a of amino acids, solving a problem that has and locked up by big drugmakers. Using
AI-first drug startups are attractive to algorithms for drug development also
diagnostic decision. vexed scientists for many years. Deep- investors. Recursion raised $121 million brings up a host of ethical questions. Will
Mind’s system, AlphaFold, will allow in 2019 before spinning off CereXis, a
scientists to synthesize new drugs to treat bias invade drug discovery much like it
new independent entity to study rare has other arenas of AI, thereby margin-
diseases and develop enzymes that might brain cancers. Nearly every major phar-
someday break down pollution. alizing certain patients or diseases? Do

algorithms need their own clinical trials? was riddled with bias when it was first the software. In a trial, an AI screening the virus might look different to the im- Protein Folding
Could AI be used to take shortcuts and developed using primarily white British system from Google Health and Deep- mune system. The approach is similar to In November 2020, DeepMind’s AI made
undermine the value of the science being patients. Researchers are training deep Mind outperformed human radiologists changing words in a sentence to change a big announcement: It successfully
done inside the laboratory? Advocates learning models instead, and finding gaps and reduced false positives in two large, its meaning. For example: “I laughed at determined a protein’s 3D shape from
say AI will make drug development and in patient care. clinically representative datasets from the the clown” versus “I cried at the clown.” its amino-acid sequence. The program,
clinical trials more efficient, thereby U.S. and U.K. By using this kind of modeling before AlphaFold, outperformed an estimated
cutting drug prices and paving the way Deep Learning Applied to mutations occur, public health officials 100 teams in a biennial protein-structure
for more personalized medicine. Medical Imaging NLP Algorithms Detect Virus could strategize and potentially prevent prediction challenge called Critical As-
Mutations new viral spreads. sessment of Structure Prediction (CASP).
Radiologists and pathologists increasing-
AI Improves Patient Outcomes ly rely on AI to assist them with diagnos- Natural language processing (NLP) algo- Predicting protein structures has long
New medical algorithms address the tic medical imaging. Last year, new U.S. rithms, which are typically used for text, Diagnostics Without Tests vexed biologists. AlphaFold had previ-
level of patient care in the U.S. Different Food and Drug Administration approvals words, and sentences, are being used The Covid Symptom Study—created by ously bested other teams, but it worked
patients experience symptoms different- allowed new products to be used widely to interpret genetic changes in viruses. doctors and scientists at Massachusetts so quickly and so accurately at last year’s
ly, and their care is based on how they in hospitals and clinics. So far, most of Protein sequences and genetic codes can General Hospital, the Harvard T.H. Chan CASP that it signaled a near future when
describe their symptoms and how those the approved devices augment (rath- be modeled using NLP techniques—and School of Public Health, King’s College the technology could be used regularly by
symptoms are interpreted by doctors. For er than fully automate) the process of can be manipulated the way you’d write London and Stanford University School other scientists.
example, assessing the severity of arthrit- reviewing images and making diagnoses. words and sentences in word processing of Medicine, and health science company
ic pain is challenging. There is a standard But emerging autonomous products are software. At MIT, computational biolo- ZOE—developed an app to study Covid Dream Communication
scoring system to rate pain, which looks making their way into clinical settings. gists used NLP to solve a vexing problem symptoms and track the spread of the Scientists discovered how to establish
at the amount of structural damage IDx-DR is an AI-enabled device that when developing new vaccines. “Viral virus. It collects and uses AI to analyze two-way communication channels
and missing cartilage, but data from detects diabetic retinopathy using retinal escape” is the ability for a virus to mutate data from 4 million global contributors between lucid dreamers. Lucid dreamers
the National Institutes of Health found images. Caption Health uses AI to cap- and evade the human immune system to discover new symptoms, predict Covid are aware that they are asleep and can
that Black patients’ pain is underscored. ture ultrasound images of the heart that and cause infection. Massachusetts Insti- hot spots, and eventually predict Covid steer their dreams. In four global studies,
It’s likely that the system itself, called expands who can read such scans. Nurses tute of Technology researchers modeled cases without physical tests. participants were outfitted with sensors
the Kellgren-Lawrence Grade (KLG), would just need a few days of training on viral escape using NLP to identify how attached to their heads and faces, and

their data was fed into a computer that case, and realtors could judge how seri-
looked for patterns. Scientists verified ous a homebuyer is. But the ethics are, of
a state of REM sleep, and then partici- course, concerning.
pants interacted with researchers using
eye movements. The study proved that
there are new ways to send and receive
real-time information while dreaming. Developing AI systems based on biological models—or deep neural networks—is
among the 10 fastest growing technologies in the U.S., as indicated by patent ap-
Thought Detection plications. The number of patents with deep neural networks grew 67% between
2016 and 2020, making it the top technology mentioned.
Deep neural networks are being used to
analyze emotional states using wireless
signals. Researchers at Queen Mary Uni-
versity of London used radio signals to Leading Applicants, 2016–2020
measure subjects as they watched a video.
The deep neural network analyzed subtle
body movements—breathing rates, heart
2,789
rate—to reveal otherwise hidden infor-
mation. If this reminds you of the “Black
Mirror” “Crocodile” episode, you’re not
far off: Research labs are developing new
1,451
AI is used to increase the speed and efficiency of new drug discovery.
technologies to read our minds. There 1,253
1,102 1,044
are business implications: HR depart-
ments could determine what employees
really think of company policies, lawyers
IBM Google Samsung Intel Microsoft
could determine how jurors lean in a
Source: IFI Claims Patent Services, January 2021 study.

Deep Twins in the OR
Mid-future optimistic scenario

The success rate of complicated medical procedures sky-
rockets when hospitals construct AI-enabled digital twins
of patients. These virtual facsimiles mirror the exterior
and interior structure of the patient’s living body, as well
as its real-time bodily functions, thanks to ultra-high
fidelity mapping technology. Doctors are thereby able to
virtually “operate” on the digital twin, using sophisticated
AI to simulate outcomes and determine the safest and
most effective approach to surgery, minimizing risk to
the patient’s life and well-being. Should a complication
arise, the algorithm adjusts and adapts the recommend-
ed course of action based on the new data—in cases
of extreme emergency, the AI can even take over and
automate certain surgical tasks. The patient experience
is revolutionized and the statistical risk of operations is
drastically reduced, making trips to the OR feel more like
a routine doctor’s visit.

Consumer
Consumer Trends
Zero UIs Consumer-grade AI Applications language processing, along with our data,
Modern interfaces are able to do more Low-code and no-code offerings from to anticipate what we want or need to do
for us with fewer direct actions—yet still Amazon Web Services (AWS), Azure, next, sometimes before we even know
captivate our attention. The average and Google Cloud will start to trickle to ask. Alibaba’s highly advanced DA
adult now makes more than 20,000 de- down to everyday people, who will create can not only interact with real humans
cisions a day—some big, such as whether their own AI applications and deploy but also deftly handle interruptions and
or not to invest in the stock market, and them as easily as they can a website. open-ended answers. Similar to Google’s
some small, such as whether to glance at We’re seeing a shift from highly techni- Duplex, Tiān Māo can make calls on your
a mobile phone when the screen lights cal AI applications used by professional behalf, but it also understands intent. So
up. Zero user interfaces—otherwise researchers to more lightweight, us- if you’re trying to schedule an appoint-
known as ambient computing systems— er-friendly apps intended for tech-savvy ment and mention that you’re usually
promise to prioritize those decisions, consumers. New automated machine commuting in the morning, the system
delegate them on our behalf, and even learning platforms make it possible for infers that you won’t be available then.
autonomously answer for us, depending nonexperts to build and deploy predictive In 2017, Future Today Institute’s model
on the circumstance. Much of this invisi- models. Platforms hope that in the near correctly projected that nearly half of
ble decision-making will happen without future, we’ll use various AI applications Americans would own and use a digital
direct supervision or input from people. as part of our daily work, just as we do assistant by 2020. (An estimated 62% of
What makes ambient design so tantaliz- Microsoft Office and Google Docs today. Americans use digital assistants today.)
ing is that it should require us to make Amazon and Google dominate the smart
fewer and fewer decisions in the near speaker market, but digital assistants can
Ubiquitous Digital Assistants be found in many places. Thousands of
future. Think of it as a sort of autocom-
plete for intent. Digital assistants (DAs)—like Siri, Alexa, applications and gadgets now track and
Replika is a programmable digital twin that you and their Chinese counterpart Tiān Māo respond to DAs. News organizations, en-
can deploy for your friends. from Alibaba—use semantic and natural tertainment companies, marketers, credit

Consumer Trends
card companies, banks, local authorities, Personal Digital Twins

political campaigns, and many others can A number of startups are building cus-
harness DAs to both surface and deliver tomizable, trainable platforms capable of
critical information. learning from you—and then represent-
ing you online via personal digital twins.
Deepfakes for Fun In 2021, China’s annual Spring Festival
Faceswap is a free and open-source deep- Gala on the country’s state broadcaster
fake app powered by TensorFlow, Keras, (CCTV) included performances from
and Python. Deep Art Effects offers synthesized celebrities. With an estimat-
desktop and mobile apps to turn images ed billion people watching, the AI copies
into stylized art. REFACE is a face swap mimicked their human counterparts
app that morphs your face onto celebrity without pre-scripted behaviors, speech-
bodies and creates GIFs to share on social es, or routines. Meanwhile, Replika is
media. Jiggy is a deepfake that makes a programmable digital twin that you
anyone dance. For now, they all result can deploy for your friends. Molly, a Y
in images and GIFs that look like they’ve Combinator–backed startup, answers
been manipulated—but with the technol- questions via text. The near future could
ogy becoming so easy to use, how long include digital twins for professionals
until we can’t tell real from fake? across a range of fields, including health
and education.
Alibaba’s voice assistant uses natural language processing.

Research
Research Trends
Closed-Source Code cation landscape. In the past four years, Most Active Institutions for AI Research
Code is important for reproducibility, Facebook seems to have gained ground.

Google
accountability, and transparency, and Of the conference papers that mention
Stanford University
it is a key to driving improvements in the framework the researchers used, 75% Carnegie Mellon University
the greater AI community. But when cited PyTorch but not TensorFlow. Of MIT
academic researchers publish papers, the 161 researchers who published more Microsoft Research
UC Berkley
they don’t often include all of their code. TensorFlow papers than PyTorch papers, Columbia University
The reason given: The code they used 55% of them switched to PyTorch, while University of Oxford
is intermingled with other proprietary only 15% moved in the other direction. Tsinghua University
Facebook
research, and it therefore can’t be re- Cornell University
leased. Fewer than 15% of all academic Cost of Training Models University of Texas at Austin
Princeton University
papers on AI publish their full code, and It costs a lot to train a model. Several UCLA
some big names—DeepMind and Ope- variables influence those costs, all of University of Illinois at Urbana-Champaign
nAI—notoriously leave theirs out, citing which have increased in the past few
INRIA
Georgia Tech
proprietary concerns. years. For example, it costs an average Peking University
of $1 per 1000 parameters. OpenAI’s IBM

University of Toronto
Framework Consolidation 175 billion parameter, GPT-3, likely University of Washington
Google’s TensorFlow and Facebook’s cost more than $10 million to train. For ETH Zurich
PyTorch are two popular frameworks smaller research groups and companies, EPFL
New York University
used by researchers, and the relative pop- the costs are out of reach. Some in the AI Duke University
ularity of different frameworks typically community are instead allowing the big
0 10 20 30 40 50 60 70 80 90
mirrors trends in the commercial appli- tech companies to pre-train and publish
big models. USA UK China Europe Canada
Papers published at NeurlIPS 2019 (fractional count)
Source: https://macropolo.org/digital-projects/the-global-ai-talent-tracker/
Research Trends
NLP Benchmarks resources, and a new public leaderboard. AI Summarizing Itself Graph Neural Networks 2016, it is a new framework that makes
The General Language Understanding We predict that by the end of 2021, this A new AI model can summarize scientific Because we perceive scents using mil- it possible for algorithms to use data
Evaluation (GLUE) benchmark is a new benchmark will also be surpassed. literature, including research about itself. lions of sensory neurons in our brains, on devices—such as mobile phones and
collection of resources for training, eval- The Allen Institute for Artificial Intelli- and because scents are multifaceted, smart watches—without compromising
uating, and analyzing natural language Machine Reading Comprehension gence (AI2) used the model in Semantic predicting the way something will smell user privacy. Research in this space has
understanding systems. It includes a Scholar, an AI-powered scientific paper is incredibly complex. For example, dramatically increased.
For AI researchers, machine reading
benchmark of nine sentence- or sen- comprehension (MRC) has been a chal- search engine to provide a short sum- how would you describe the smell of an
tence-pair language-understanding tasks lenging goal, but an important one. MRC mary of papers on AI. What makes this orange? Sweet? Bright? Grassy? Each GP Models
built on existing datasets and selected makes it possible for systems to read, work impressive—and ultimately so use- descriptor is unique. Classifying smell is Gaussian processes (GP) are the gold
to cover a diverse range of dataset sizes, infer meaning, and immediately deliver ful—is that it is capable of compressing tricky because it requires a multi-label standard for many real-world modeling
text genres, and degrees of difficulty. It answers while sifting through enor- long papers with accuracy and efficiency. system. Graph neural networks (GNNs) problems, especially in cases where a
includes a diagnostic dataset designed to mous datasets. In 2019, China’s Alibaba constitute a particular type of deep model’s success hinges on its ability to
evaluate and analyze model performance outperformed humans when tested by No Retraining Required neural network that operates on graphs faithfully represent predictive uncertain-
with respect to a wide range of linguistic the Microsoft Machine Reading Com- as inputs. GNNs are being used to detect ty. GPs are becoming more accurate and
phenomena found in natural language. Training robots to do more than one smell—to predict odors at a molecular
prehension dataset (or MS MARCO for thing is difficult, but a new model pits easier to train, benefiting from neural
And it includes a public leaderboard so short), which assessed its ability to use level—and for a wide array of chemical network improvements.
that researchers can track their perfor- identical robot arms against one another and biological processes. For example,
natural language to answer real questions in a game (moving objects on a virtual
mance. The human baseline score is 87, posed by humans. Alibaba’s system deliv- researchers at the Broad Institute used
and between May 2018 and August 2020, table in specific ways) in which one robot them to discover antibiotic compounds GPT-3’s Influence
ered answers to search queries posted by challenges the other with increasingly
natural language processing systems people to Microsoft’s Bing, such as “How that don’t have toxic side effects. The enormous AI that generates hu-
increased from 60 to 90.6, surpassing difficult tasks. It’s an example of multi- man-like language, GPT-3, was released
many carbs are in an English muffin?” task learning, a deep learning model in
humans. The SuperGLUE benchmark and “How do you grow hops?” Federated Learning by OpenAI last year. The text generator
is a new measurement of more difficult which machines learn different skills as has written blog posts and code. It was
language understanding tasks, improved they progress. OpenAI’s model allows Federated learning is a technique that pitted against college students in an essay
a bot to solve new kinds of problems distributes machine learning to the edge. writing contest, and the anonymized
without requiring retraining. Introduced by Google researchers in

Research Trends
papers were graded by professors. GPT-3 language-only data by contextually deeply culturally embedded racism and
earned mostly B’s—the same as human mapping language “tokens,” or the words sexism. A few years ago, if you typed
students. But the AI has demonstrated used to train language models, to related “CEO” into Google Images, the first
a strong anti-Muslim bias. Researchers images, or “vokens.” For example, au- result of a woman was CEO Barbie. In
from Stanford University and McMaster to-generated image captions often can’t an experiment, researchers at Carnegie
University probed the neural network infer context. Vokenization would enable Mellon University trained a system to
on tasks including prompt completion, machines not just to recognize objects autocomplete images of men and wom-
analogical reasoning, and story genera- but to really “see” what’s in them. en cropped below the neck. In pictures
tion. They found that a Muslim-violence of men, the system autocompleted him
bias appears consistently and creatively Machine Image Completion wearing a suit. The system autocomplet-
in many use cases of the model. It’s yet ed women—including U.S. Rep. Alexan-
another example of how bias creeps into If a computer system has access to dria Ocasio-Cortez (D-N.Y.)—wearing a
our automated systems. Left unchecked, enough images—say, millions and low-cut top or bikini 53% of the time.
it will cause problems throughout society millions—it can patch and fill in holes
as AI matures. in pictures. This capability has practical
applications for professional photog- Predictive Models Using Single
raphers, as well as for everyone who Images
Vokenization wants to take a better selfie. Soon, if Computer vision systems are getting
Models like GPT-3 are trained on syntax the foreground of a mountain is out of smarter. Neural networks can predict
and grammar, not creativity or common focus, or if your skin has an unsightly geometry from a single color image. In
sense. So researchers at the University of blemish, another version can be swapped 2019, the DeepMind team developed a
The SuperGLUE benchmark will be broken by the end of 2021.
North Carolina–Chapel Hill are com- in to generate the perfect picture. As generative adversarial network (GAN)
bining language models with computer such technology becomes commonplace, that creates videos from images. For
vision. Humans learn in a multilayered, there will be significant biases and other example: Imagine a photo of a person
multidimensional way, so a new tech- pitfalls to navigate. For example, image holding a basketball. Based on his pos-
nique called vokenization extrapolates generation algorithms routinely reflect ture, face, and other data within the

Research Trends
picture, the GAN figures out what likely Model-free Approaches to RL a $10 million grant program to catalyze The U.S. Army Research Laboratory Classic AI is the former, because it more
happened next and generates a video Dreamer is a reinforcement learning research in this area, although all of the has a system that uses a brain-computer closely represents how we understand
clip of the action. Earlier, researchers at (RL) agent that uses a world model to big tech companies are working closely interface armed with computer vision human thought—and the original intent
MIT’s Computer Science and Artificial learn long-sighted predictions, employ- to advance RTML too. technology and allows a person to rapidly was to teach machines to think like us.
Intelligence Laboratory (CSAIL) trained ing backpropagation through model see and sort images within her line of Researchers are working on new ways to
computers to predict what humans predictions. It can create models from Automated Machine Learning sight. CloudSight, a technology company combine both learning and logic using
would do next using YouTube videos raw images and learn from thousands (AutoML) specializing in image captioning, is work- neural networks, which would under-
and TV shows such as “The Office” and of predicted sequences in parallel using ing on a hybrid crowdsourced computer stand data through symbols rather than
“Desperate Housewives.” CSAIL’s system Some organizations want to move away vision system. Microsoft researchers always relying on human programmers
a graphics processing unit (GPU). This from traditional machine learning meth-
predicts whether two people are likely to new approach solves long-horizon tasks have proposed Pandora, a set of hybrid to sort, tag, and catalog data for them.
hug, kiss, shake hands, or slap a high five. ods, which are time-consuming and diffi- human-machine methods and tools for Symbolic algorithms will aid the process,
using an imagined world. cult and require data scientists, specialists
SinGAN is an unconditional generative understanding system failures. Pandora which should eventually lead to robust
scheme that can manipulate and enhance in AI fields, and engineers. Automated leverages both human and system-gen- systems that don’t always require a hu-
images—sketch a mountain, and it will Real-time Machine Learning machine learning, or AutoML, is a new erated observations to explain malfunc- man for training.
produce a realistic-looking synthetic One big challenge in AI is building approach: a process in which raw data tions related to input content and system
photograph. This research will someday machines that can proactively collect and and models are matched together to re- architecture.
veal the most relevant information. Goo- General Reinforcement Learning
enable robots to more easily navigate interpret data, spot patterns and incor- Algorithms
human environments—and to interact porate context, and ultimately learn in gle, Amazon, and Microsoft now offer a
host of AutoML products and services. Neuro-Symbolic AI Researchers are developing single
with us humans by taking cues from our real time. New research into real-time
body language. Retail, manufacturing, machine learning (RTML) shows that The development of AI has been on algorithms that can learn multiple tasks.
and education settings could be especially it’s possible to use a continual flow of Hybrid Human-Computer Vision two conceptual tracks since the 1950s: DeepMind, the team behind AlphaGo,
relevant. data and adjust models in real time. This symbolic (machines that use a base of which learned how to play Go with the
AI isn’t yet capable of fully functioning knowledge and rules that represent skill level of a human grandmaster, has
signals a big change in how data moves, without human assistance. Hybrid in-
and in how we retrieve information. The concepts) and non-symbolic (machines developed an innovative new algorithm:
telligence systems combine humans and that use raw data to create their own pat- AlphaZero. It is capable of achieving
National Science Foundation launched AI systems to achieve greater accuracy. terns and representations of concepts). superhuman performance not only in

Research Trends
Go but in other games as well, including tonomous and incremental skill building independently on different algorithms wrote in Python. It’s a move that signals
chess and shogi (Japanese chess). This and development, and researchers will and datasets, and they only see one likely fragmentation in the future of the
one algorithm starts with no knowledge continue to push the limits of what’s another’s work once it is deployed. This AI ecosystem, not unlike the current
except for the rules of the game and possible in this field. has been the cause of recent stock market iOS/Android rivalry or the long Mac/PC
eventually develops its own strategies glitches and e-commerce website wonki- war. Businesses will find it increasingly
to beat other players. In January 2020, Proliferation of Franken- ness. It is especially challenging for big cost-prohibitive and difficult to switch
DeepMind published new research Algorithms companies like Facebook, which have between AI frameworks and languages.
showing how reinforcement learning billions of algorithms working together
techniques could be used to improve Algorithms are simply rules that define at any given time.
our understanding of mental health and and automate the treatment of data. They
motivation. are built using “if this, then that” logic
Using AI, researchers automated the task of convert-
that a computer can understand and pro- Proprietary, Homegrown AI
ing live actor performances (left) to computer game
cess. Here’s an easy example: If a website Languages
virtual characters (right).
Continuous Learning reader’s IP address is based in Baltimore, Python is a leading language with lots of
At the moment, deep learning tech- the rules then allow that reader to freely pre-built libraries and frameworks. Julia,
niques are helping systems learn to solve access the site; if the IP address is based a language developed by Massachusetts
complex tasks in a way that resembles in Belgium, then the rules first show a Institute of Technology, is an open-
what humans can do—but those tasks GDPR screen stating privacy and cookie source language that focuses on numer-
are still specific, such as beating a human policies. While a single algorithm might ical computing. And of course there’s
at a game. And they require a rigid be easily described and deployed as ex- Lisp, created by modern AI’s foreparent
sequence: Gather data, determine the pected, systems of algorithms all working John McCarthy in 1958. Companies are
goal, deploy an algorithm. This process together can sometimes pose problems. starting to build and release their own
requires humans and can be time-con- Developers don’t always know in ad- software packages now, as well as unique
suming, especially during early phases vance how one algorithm will function programming languages for AI applica-
when supervised training is required. alongside other algorithms. Sometimes, tions. Uber released its own probabilistic
Continuous learning is more about au- several teams of developers are working programming language, Pyro, which it

Talent
Talent Trends
AI Brain Drain like New York University professor Yann

The brain drain of AI researchers out LeCun and Carnegie Mellon professor Enrollment in Stanford University’s natural language processing class is
of academia and into corporations is Jessica Hodgins. Poaching departments now 10 times the size it was in 2004.
growing at an alarming pace. The reason today could rob the future of future AI
is simple: compensation packages. Top experts: Without great scholars, who will
academics earn generous salaries and train the next generation of innovators?
benefits, and they get to work in a similar
tenured environment that’s carefully cul- AI Universities Stanford NLP class enrollment
tivated to represent their experience in Interestingly, new institutions solely 1200
academia. Google, DeepMind, Amazon, focused on AI are launching around
and Microsoft hired 52 tenured and ten- the world and offering postsecondary
CS224D (DL NLP)
1000
ure-track professors from U.S. univer- education. In the United Arab Emirates,
CS224S (speech)
sities between 2004 and 2018. In return the Mohamed bin Zayed University
CS276 (IR)
800
for their poaching, tech companies are of Artificial Intelligence launched last
CS224U (NLU)
endowing AI professorships at top uni- year. MBZUAI is the world’s first grad- 600
CS124 (ugrad HLT)
versities. In some cases, professors take uate-level, research-based AI university.

CS224N (NLP)
one- or two-year sabbaticals to work at Based in Abu Dhabi, it offers master of 400
tech companies and then return to their science in computer vision and Ph.D.
universities—but corporate benefits can programs in AI-related fields. Founded 200
be difficult to give up. In one infamous by Harvard University and University
case, Uber poached an entire robotics lab of California–Los Angeles faculty, Univ. 0
from Carnegie Mellon University—40 AI is an online program for training in
1998-1999
1999-2000
2000-2001
2001-2002
2002-2003
2003-2004
2004-2005
2005-2006
2006-2007
2007-2008
2008-2009
2009-2010
2010-2011
2011-2012
2012-2013
2013-2014
2014-2015
2015-2016
2016-2017
2017-2018
2018-2019
The Mohamed bin Zayed University of AI opened professors and researchers in total. Face-
in the United Arab Emirates last year. machine learning and AI.
book has aggressively hired superstars
Source: https://nlp.stanford.edu/

Talent Trends
I have always been Demand for AI Talent Growing Corporate AI Labs AI for Interviews algorithms work invisibly, which means
For many years, demand for AI talent AI labs are located around the world, Recognition systems can now be de- that this is an area that could face regula-
convinced that has outpaced supply. In the U.S., there with concentrations in North America, ployed to watch you being interviewed tory scrutiny. Consumer advocacy orga-
nization Electronic Privacy Information
the only way to get were nearly three times more AI-related Europe, and Asia. Facebook, Google, and to gauge your enthusiasm, tenacity,
job postings on Indeed last year than IBM, and Microsoft operate 62 labs and poise. Algorithms analyze hundreds Center filed a complaint with the U.S.
artificial intelligence job views for AI-related roles. While dedicated to AI R&D, and the majority of details, such as the tone of your voice, Federal Trade Commission requesting
an investigation into HireVue, alleging
to work is to do the schools are adding programs, increasing are outside of the U.S. because of access your facial expressions, and your man-
its tools produce results that are “biased,
enrollment, and adding classes, there are to talent. During the Trump administra- nerisms to best predict how you’ll fit in
computation in a way just too many new needs for AI skills and tion, immigration restrictions and strin- with the culture of a community. Start- unprovable, and not replicable” through
algorithmic models.
similar to the human nowhere near enough trained workers. gent visa requirements made it difficult ups such as HireVue use AI systems to
As demand grows, the hiring process is to recruit talent into the United States, help companies decide which candidates
brain. That is the goal taking longer and becoming more expen- and overseas labs allowed companies to to hire. But this kind of recognition tech-
I have been pursuing. sive. This is impeding growth at some overcome that barrier. Most of those labs nology has practical applications well be-
companies, according to a 2020 Reuters do basic AI research rather than product yond job interviews: It can detect when
We are making study. Demand has also driven up wages. development. The G-MAFIA—Google, someone is likely to make a purchase—or
progress, though we A recent study from Glassdoor estimated Microsoft, Amazon, Facebook, IBM, attempt to shoplift—in a store, whether
that average annual salaries for AI-relat- and Apple—spend $76 billion on R&D someone is lying, and whether someone
still have lots to learn ed jobs rose 11% between 2017 and 2018 annually. Collectively, their market cap is receptive to new suggestions and ideas.
to $123,069. exceeds $6 trillion. Unlike security cameras, which tend to
about how the brain
have a light indicating they’re recording,
actually works.”
— Geoffrey Hinton

Creative
Creative Trends
Assisted Creativity Generative Algorithms for Generating Virtual Environments

Generative adversarial networks (GANs) Content Production from Short Videos
are capable of far more than generating Last year, the creators of “South Park” Chip designer Nvidia is teaching AI to
deepfake videos. Researchers are partner- built an entire show using deepfakes. build realistic 3D environments from
ing with artists and musicians to generate “Sassy Justice” aired on YouTube and short video clips. The method builds
entirely new forms of creative expres- featured a synthetic reporter, Fred Sassy, on previous research on GANs. Nvid-
sion. From synthesizing African tribal who looked a lot like former President ia’s system generated graphics based on
masks to building fantastical, fictional Trump—with just a different enough open-source datasets used by the auton-
galaxies, AI is being used to explore voice and hairstyle to evade legal chal- omous driving field. Using short clips
new ideas. Last year, Nvidia launched lenges. Episodes featured deepfakes of Al segmented into various categories—such
GauGAN (named after post-Impres- Gore, Mark Zuckerberg, Jared Kushner, as buildings, sky, vehicles, signs, trees, or
sionist painter Paul Gauguin), a gen- and others. The open-source algorithm people—the GANs created new, different
erative adversarial AI system that lets DeepFaceLab has been used by other versions of these objects. The array of
users create lifelike landscape images artists and filmmakers. There was a Hulu possible applications is vast. Automat-
that never existed. The National Insti- commercial deepfaking sports stars, and ically generated virtual environments
tute of Informatics in Tokyo built an several 2021 Super Bowl commercials, could be used for fantasy and superhero
AI lyricist, while Amazon released its including an eerie appearance by Vince movies and could bring down the costs of
DeepComposer system, which composes Lombardi, used deepfakes and synthetic TV production and game development.
music “automagically.” These AIs are not media. OpenAI’s deep learning algorithm
ostensibly intended to replace artists, but released a neural network called Juke-
Creative studio SoKrispyMedia develops short films featuring stick figures in battle. It relies on real-time rather to enhance their creative process. box that generates songs in a bunch of
rendering for photorealistic results. different styles and simulated voices that
Image credit: Nvidia. sound (sort of) like Elvis and others.

Creative Trends
Automated Versioning Automatic Voice Cloning and Automatic Ambient Noise

Journalists at Switzerland-based Tamedia Dubbing Dubbing
experimented with generative techniques Anyone who’s ever recorded a podcast is For some time, we’ve been training com-
during their country’s 2018 election. A familiar with editing challenges such as puters to watch videos and predict corre-
decision-tree algorithm Tamedia named guests talking over each other, interrup- sponding sounds in our physical world.
Tobi generated automated articles de- tions from sirens and other background For example, what sound is generated
tailing vote results for each municipality noises, and inconvenient sneezes. Those when a wooden drumstick taps a couch?
covered by the private media group’s 30 moments can stop a conversation cold. A pile of leaves? A glass windowpane?
newspapers. It also produced content But what if you could edit the spoken The focus of this research, underway at
simultaneously in multiple languages. word the way you edit a word document? MIT’s Computer Science and Artificial
In total, Tobi published 39,996 different That’s the promise of AI companies in- Intelligence Laboratory, should help
versions of election stories that averaged cluding Resemble AI and Descript, which systems understand how objects interact
250 words each. The articles carried a make it possible to clone voices. That with each other in the physical realm.
special byline alerting readers that they’d means soon you might see a star like Numerous projects are now underway to
been written by an algorithm. With Phoebe Waller-Bridge in a movie and make it easier to automatically generate
more experiments underway, we expect also hear her, in her own voice, speaking voices, videos, and even storylines.
to see news and entertainment media in Portuguese. There’s obviously a dark
companies developing multiple versions side to this technology, however. Last
of the same content to reach wider audi- year, hackers used voice cloning tools to
ences or to produce massive amounts of trick an employee into thinking he was
Jukebox is a neural net that generates music, including rudimentary singing, as raw audio in a variety of content at scale. speaking on the phone to his CEO; he
genres and artist styles. then transferred $243,000 to a scammer’s
bank account.

Geopolitics
& Defense
Geopolitics and Defense Trends
AI Nationalism National Science Foundation with $100 zens; and the United Arab Emirates has
Governments are instituting new re- billion in funding over five years and ear- a sweeping set of policy initiatives on AI
strictions on mergers and acquisitions marks $10 billion for local and regional and appointed Omar Sultan Al Olama as
and investment activity to ensure that tech hubs across America. Meanwhile, in its minister of state for artificial intelli-
AI developed by companies does not aid China the Ministry of Science and Tech- gence. In the U.S., numerous public and
foreign adversaries. In 2020, Germany nology established 20-city AI pilot zones private groups work independently on
passed a law to allow the government to that should open by 2023. They will carry the future of AI on behalf of the nation.
review or block investments or takeovers out AI-based policy experiments and Those efforts, however, lack interagency
of robotics, AI, and semiconductor com- social experiments, according to official collaboration and coordinated efforts to
panies by non-EU–based firms. The for- government documents. streamline goals, outcomes, R&D efforts,
eign ownership threshold also dropped and funding.
to 10% from 25%. The U.K. similarly National AI Strategies
strengthened its foreign investment pol- A new wave of countries will launch AI as Critical Infrastructure
icies for AI: Any M&A activity involving national AI strategies in 2021 and 2022. Government researchers are exploring
AI companies in which revenue exceeds Iceland, Norway, Kenya, Chile, Colom- ways to spearhead AI development for
1 million pounds requires government bia, and New Zealand are all developing critical systems use: road and rail trans-
scrutiny. Many acquisitions are being national strategies. China passed its New portation systems; power generation and
blocked as a result. The U.K. and U.S. Generation Artificial Intelligence Devel- distribution; and predicting routes for
blocked a takeover of Imagination Tech- opment Plan with aggressive benchmarks public safety vehicles, such as ambulances
nologies by China’s Canyon Bridge last to become the world’s dominant AI and firetrucks. Rather than shunning AI
Cyber warfare will change the art of war in the coming decade. year. A bipartisan bill proposed in the player within 10 years; France adopted a systems, there is new interest in using
Senate, known as the Endless Frontier national strategy called AI for Humanity; the technology to prevent disasters and
Act, explicitly frames AI as a race be- Saudi Arabia has both a strategy and a improve safety.
tween the U.S. and China. It also creates legal framework for making robots citi-
a new technology directorate within the

Nation-based Guardrails and systems interoperable. The EU and the C felony. Bills have also been introduced ity of output. It’s plausible that various The New Mil-Tech Industrial
Regulations Organisation for Economic Co-operation in California, Texas, and Massachusetts, countries will enact new regulations Complex
From self-driving car accidents to and Development similarly published and a number of federal bills are being requiring explainability in the coming In the past few years, some of the biggest
election interference through disinfor- their own guidelines, and the Trump discussed. These initiatives will likely years. Imagine sitting beside a genius AI companies in the U.S. have partnered
mation campaigns to political repres- administration signed an executive order be met with arguments that prohibiting mathematician who gives you correct an- with the military to advance R&D and
sion enhanced by facial recognition and to spur the development of standards deepfakes infringes on free speech rights. swers in Italy, but receiving her answers find efficiencies. In fact, the public sector
automated surveillance, major events in the U.S. While these efforts could across the border in France would mean cannot advance its technology without
over the past few years have thrown introduce new ways to safeguard against Making AI Explain Itself asking her to stop and show her work— help from outside companies. Plus, there
into sharp relief the dangers of artificial bias and to ensure trust, they also each and every time she’s asked to share her is a lot of money to be made. The U.S.
attempt to create strategic advantages for You’ve undoubtedly heard someone answers in a new country.
intelligence. Few guardrails now exist for argue that AI is becoming a “black box”— General Services Administration and the
a technology that will touch every facet stakeholders. As AI continues to develop Department of Defense’s Joint Artificial
according to different rules in China, the that even researchers working in the
of humanity, and countries are racing field don’t understand how our newest New Strategic Technical Alliances Intelligence Center recently awarded a
to develop and publish their own AI EU, and the U.S., one of the hallmarks five-year, $800 million contract to Booz
of the field—global academic collabora- systems work. That’s not entirely true. New strategic technical alliances between
strategies and guidelines. The European However, there is growing concern countries will help drive future R&D but Allen Hamilton for AI product develop-
Union developed an AI Alliance and plan tion—could drastically decline. ment. The U.S. Army awarded Lockheed
among computer scientists, journalists, could also strain existing geopolitical al-
of cooperation between member coun- and legal scholars who argue that AI liances or heighten tensions. Likely part- Martin a $75 million contract for a ma-
tries, and Estonia is developing its own Regulating Deepfakes systems shouldn’t be so secretive, and ners include the U.S., Germany, Japan, chine learning cyber jamming pod that
legal framework governing the use of New measures to regulate the creation regulators are paying close attention. India, South Korea, the U.K., France, can be mounted on Humvees or drones.
AI within the country. Last year, China and distribution of deepfakes will be Broadly speaking, a few challenges must and Canada—leaving China and Russia With a new focus on defense roadmaps
moved into position to lead the first introduced throughout 2021 in the U.S. be overcome. Requiring transparency in to partner up separately. The latter two that include AI components, startups
set of global AI norms and standards. and elsewhere. A bill in Hawaii’s state AI could reveal a company’s trade secrets. countries have already announced a tech- working in high-resolution satellite im-
In 2019, the country published a report legislature seeks to prohibit unautho- Asking the systems to explain their nical alliance on satellites and deep-space agery, computer vision, and unmanned
on technical standards that would allow rized deepfake apps and tools. If it passes, decision-making processes as they work exploration. aerial vehicles are attracting lucrative
companies to collaborate and make their deepfaking would be considered a Class could also degrade the speed and qual- venture capital investment. Both Am-
azon and Microsoft made headlines

over a $10 billion, 10-year government dozens resigned. Eventually, Google said airstrikes. Future Today Institute analysis
tech contract called the Joint Enterprise it wouldn’t renew its contract on the shows that the future of warfare encom-
Defense Infrastructure, or JEDI. Others, project. The company launched a set of passes more than traditional weapons.
including IBM, Oracle, and Google, also ethical principles governing its develop- Using AI techniques, a military can
competed to transform the military’s ment and use of AI, including a provision “win” by destabilizing an economy rather
cloud computing systems. Meanwhile, that prohibits any systems from being than demolishing countrysides and city
the CIA awarded Amazon a $600 million used for “weapons or other technologies centers. From that perspective, China’s
cloud services contract, while Microsoft whose principal purpose or implementa- unified march to advance AI puts the
won a $480 million contract to build tion is to cause or directly facilitate injury emerging superpower dangerously far
HoloLens headsets for the Army. The to people.” ahead of the West.
contracts prompted employee protests.
In 2017, the Department of Defense Algorithmic Warfighting
established an Algorithmic Warfare
Cross-Functional Team to work on Proj- Future wars will be fought in code, using
ect Maven—a computer vision and deep data and algorithms as powerful weap-
learning system that recognizes objects ons. The current global order is being
from still images and videos. The De- shaped by artificial intelligence, and the
partment of Defense contracted Google same countries leading the world in AI
to train the AI systems to analyze drone research—the U.S., China, Israel, France,
footage, but it turned out the Googlers Russia, the U.K., and South Korea—are
assigned to the project didn’t know they also developing weapons systems that
were doing military work. A high-pro- include at least some autonomous func-
file backlash ensued: As many as 4,000 tionality. Israel uses autonomous drones
Google employees signed a petition ob- for border patrol, while China developed
U.S. Rep. Ro Khanna (D-Calif., pictured here), along with U.S. Sens. Todd Young (R-Ind.) and Chuck Schumer
jecting to Project Maven, and ultimately stealth drones capable of autonomous
(D-N.Y.), and U.S. Rep. Mike Gallagher (R-Wis.), unveiled the bipartisan, bicameral Endless Frontier Act.

China’s AI Rules
If you think of China as a country startup SenseTime is pioneering myriad WeChat) began hunting for American
that copies rather than innovates, recognition technologies, such as a sys- talent when it opened an AI lab in Seattle
think again. tem that provides advertisers real-time three years ago. It has since upped its
feedback on what people are watching; stakes in companies including Tesla and
China is a global leader in artificial intel-
technology that can extract customer Snap. The payoff for the Chinese is not
ligence. Under President Xi Jinping, the
information and carry out statistical just a typical return on investment—Chi-
country has made tremendous strides in
analysis in crowded areas like shopping nese firms expect IP as well. China-based
many fields, but especially in AI. Busi-
malls and supermarkets; and simultane- AI startups now account for nearly half
nesses and the government have collabo-
ous recognition of everything in a scene, of all AI investments globally.
rated on a sweeping plan to make China
including people, pets, automobiles,
the world’s primary AI innovation center
trees, or soda cans. Strategic panopticon
by 2030, and it’s already making serious
progress toward that goal. That plan is That gives China an incredible advan- In late 2019, China began requiring all
unlikely to be repealed by a new govern- tage over the West. It also gives three of citizens to submit to facial recognition in
ment; China abolished Xi’s term limits China’s biggest companies—Baidu, Alib- order to apply for new internet or mobile
and will effectively allow him to remain aba, and Tencent—superpowers. Collec- services, and began requiring that tele-
in power for life. tively, they’re known as the BAT, and com companies deploy AI to check the
they’re all part of the country’s well-capi- identities of people registering SIM cards.
The country’s enormous population of
talized, highly organized AI plan. Chinese social media platforms require
1.4 billion offers researchers and start-
The BAT is important to you even if users to sign up with their real names.
ups there a command of what may be
you’ve never used them and don’t do any In Chinese schools, surveillance cameras
the most valuable natural resource in
President Xi Jinping has foregrounded AI as a major, long-term priority in China. business in China. That’s because these with computer vision are used widely
the future—human data—without the
companies are now well established in and track whether students are paying
privacy and security restrictions common
Seattle and around San Francisco, and attention and whether they attempt to
in much of the rest of the world. If data
they are investing significantly in U.S. cheat or sleep. These and other national
is the new oil, then China is the new
startups. Baidu (a Chinese search-engine standards make it easier for the govern-
OPEC. The kind of rich data the Chinese
company often likened to Google) estab- ment to track its citizens. China’s social
are mining can be used to train AI to
lished AI research centers in Silicon Val- credit system, an algorithmic reputation
detect patterns used in everything from
ley and Seattle, and Tencent (the devel- system developed by the government,
education to manufacturing to retail
oper of the mega-popular messaging app standardizes assessments of citizens’ and
to military applications. The Chinese
businesses’ behavior and activity.

China’s AI Rules
In 2020, numerous reports of abuse a much more lucrative deal: Ecuador
revealed that China turned its AI on eventually signed away big portions of
the ethnic Uighur Muslim community. its oil reserves to China to help finance
Chinese-educated Researchers Dominate NeurIPS
Huawei developed special AI software to infrastructure projects. Similar package
identify Uighurs and alert local police. In deals have been brokered in Venezuela Chinese-educated researchers dominated the papers accepted at NeurIPS, a prestigious international AI confer-
2021, China blocked social media plat- and Bolivia. ence, according to a new study from MacroPolo, a think tank based at the Paulson Institute, which promotes con-
form Clubhouse after an open, democrat- China is quietly weaponizing AI, too. structive collaborations between the U.S. and China. Nearly one-third of the papers accepted were from China—
ic debate flourished on the platform about China’s People’s Liberation Army is more than from any other country.
the plight of the Uighur community. catching up to the U.S. military, using AI
for such tasks as spotting hidden imag-
Risk profile es with drones. The Chinese military
We have failed—and we are continuing is equipping helicopters and jet fight-
to fail—to see China as a military, eco- ers with AI. The government created
nomic, and diplomatic threat when it a top-secret military lab—a Chinese
comes to AI. China has already used its version of DARPA—and it’s building
Belt and Road Initiative as a platform to billion-dollar AI national laboratories.
build international partnerships in both China’s military is achieving remarkable
physical and digital infrastructure, and AI successes, including a recent test of
14% 12% 11% 14% 13% 17% 14% 19% 19% 24% 29%
it is making surveillance technologies “swarm intelligence” that can automate
available to countries with authoritarian dozens of armed drones. 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
regimes. Two Chinese companies —the
state-controlled CEIEC and Huawei— The bottom line
built Ecuador’s surveillance system, called There is a Chinese proverb that roughly Source: https://macropolo.org/digital-projects/the-global-ai-talent-tracker/
ECU-911. The system promised to curb translates to “forewarned is forearmed.”
high murder rates and drug crime, but Now that you know what’s coming,
it was too expensive an investment. As reframe your thinking of China as simply
a result, a deal was struck for a Chi- the world’s factory.
nese-built surveillance system financed
with Chinese loans. It was a prelude to

Society
Society Trends
Ethics Clash content.) In January 2021, the board companies working on sensitive projects. sense for the greater good and whether
On Dec. 2, 2020, Timnit Gebru, the co- made its first rulings on disputed content, Those working in information security competition will eventually be hindered
lead of Google’s ethical AI team, posted overturning four out of the five cases it and risk management should pay special (along with access), as we’ve seen in oth-
a tweet saying she’d been fired. Known saw. But there are billions of posts on attention to advances in computer vision. er fields such as telecom and cable.
for groundbreaking research in bias and Facebook every day and an untold num-
facial recognition, she is widely respect- ber of content complaints—which means Marketplace Consolidation Fragmentation
ed within the broader AI community. the oversight board operates at the speed
of traditional government. We anticipate As much as the AI ecosystem booms, a The AI ecosystem spans hundreds of
While the incident concerned a paper rush of acquisitions means consolidation, companies. They are building the net-
she co-authored and a decision by Google many more ethics clashes in 2021.
too. Big companies now snap up startups work infrastructure, the custom chipsets,
that it didn’t meet “our bar for publica- long before they have time to mature— the consumer applications, the back-end
tion,” she and others argued that Google’s Ambient Surveillance the average age at acquisition is 3 years communications systems, the low-power
ethics team was motivated by PR rather What happens behind closed doors may old. Just nine big companies dominate radios in our smart home gadgets … we
than progress. It set off a firestorm of not be secret for long, and executives the AI landscape: Google, Amazon, Mic- could go on. Meanwhile, a large number
criticism, and by February 2021 Google should beware new ambient surveillance rosoft, IBM, Facebook, and Apple in the of policy groups, advocacy organiza-
said it would change its diversity and methods. Scientists at Massachusetts U.S., and Baidu, Alibaba, and Tencent in tions, and governments are developing
research policies, “streamline its process Institute of Technology discovered how China, with significant fortification and guidelines, norms and standards, and
for publishing research,” and change how to use computer vision to track data from support from their country’s govern- policy frameworks hoping to guide the
sensitive employee exits are managed. what they call “accidental cameras.” Win- ment. On the investment side, Qual- future development of AI. As a result,
Google’s head of AI, Jeff Dean, expressed dows, mirrors, corners, houseplants, and comm, Tencent, Intel Capital, Google the ecosystem is fragmented in two ways:
his regret in an email to staff. To deal other common objects can be used, along Ventures, Nvidia, Salesforce, Samsung infrastructure standards and governance.
with its own ethical failures, Facebook with AI, to track subtle changes in light, Ventures, Alibaba, Apple, Baidu, Citi,
launched an independent oversight board shadows, and vibrations. The result: We and In-Q-Tel fund much of the growth.
Timnit Gebru is a pioneering researcher on algo- with the power to overrule content mod-
rithmic bias. all may soon have X-ray vision capabil- When it comes to the future of AI, we
eration guidelines—and even to overrule ities—which may not be great news for should ask whether consolidation makes
Image credit: Wikimedia Commons.
Mark Zuckerberg himself. (But just on

00 01 02 03 04 05 06 07 08 09 10 11 12
Expert Insight
AI reveals our real-

world biases The biggest topic of
conversation in AI right
Meredith Broussard now is bias.
Associate Professor,
New York University algorithmic bias, audit your systems
and your algorithms for bias, and re-
quire your vendors to do the same. If
The biggest topic of conversation an AI system discriminates or makes
in AI right now is bias. We’re moving decisions that are not in line with
away from the stage of marveling your corporate values, don’t use it.
at AI technology, and people are
**
realizing that AI models and auto-
mated systems reproduce all of the Meredith Broussard is an associate
existing biases and inequalities of professor at the Arthur L. Carter
the real world. Journalism Institute of New York
If your company is using AI, ask: University, an Affiliate with the New
how might this system fail, and for York University Alliance for Public In-
whom? If you are in a heavily-reg- terest Technology, and the author of
ulated industry, AI is a compliance Artificial Unintelligence: How Com-
risk. Educate your teams about puters Misunderstand the World.
© 2021 Future Today Institute

Society Trends
AI Still Has a Bias Problem forcement predict whether we’re likely to darker skin types because of how mela- experiment was to prove that AI isn’t during the hiring process. Drexel Uni-
It’s no secret AI has a serious and multi- commit a crime. Researchers at a number nin absorbs green light. This can lead to inherently biased, but that data input versity researchers built an app that uses
faceted bias problem. Just one example: of universities—including the University bias when insurance company algorithms methods—and the people inputting that biometrics to predict when dieters are
The datasets used for training often come of Maryland; Columbia University; Car- take into account heart rates, blood data—can significantly alter an AI’s be- likely to stray from their prescribed reg-
from places like Reddit, Amazon reviews, negie Mellon; MIT, Princeton Univer- pressure, and risk rates for conditions havior. In 2019, new pre-trained systems imens. Researchers at the University of
and Wikipedia, a site inherently riddled sity; University of California, Berkeley; like irregular heartbeats or a potential built for natural language generation Copenhagen created a machine learning
with bias. The people building models International Computer Science Insti- heart attack. were released—but the conversations system to spot cheating on essays with,
tend to be homogeneous and aren’t often tute; among others—are studying the side from which they learned were scraped they say, a 90% accuracy rate.
aware of their own biases. As computer effects of automatic decision-making. Problematic Training Data from Reddit and Amazon reviews, both
systems get better at making decisions, You, or someone you know, could wind of whose author populations skew white Algorithms Targeting Vulnerable
up on the wrong side of the algorithm In 2018, researchers at MIT developed and male, which means that their use of
algorithms may sort each of us into an AI called Norman that was trained Populations
groups that don’t make any obvious sense and discover you’re ineligible for a loan, language isn’t representative of everyone.
or a particular medication, or the ability to perform image captioning. They This illustrates an ongoing challenge There is no question that machine
to us—but could have massive repercus- trained Norman using only content from learning systems trained correctly can
sions. to rent an apartment, for reasons that within the developer community. It is
aren’t transparent or easy to understand. a subreddit that’s known for graphic already difficult to get authentic data help find missing children and detect
Every single day, you are creating un- violence. When Norman was ready, from real people to train systems, and abuse. The problem is that the systems
imaginable amounts of data, both actively Increasingly, data is being harvested they unleashed him against a similar with new privacy restrictions, developers use data from vulnerable populations to
(such as when uploading and tagging and sold to third parties without your neural network that had been trained are choosing to rely more on public—and do their training. The Multiple Encoun-
photos on Facebook) and passively knowledge. These biases can reinforce using standard data. Researchers fed problematic—datasets. ter Dataset contains two large datasets
(driving to work, for example). That themselves over time. As AI applications both systems Rorschach inkblots and of photos: people who have not yet
data is mined and used, often without become more ubiquitous, the negative asked them to caption what they saw, committed a crime, and an FBI dataset
your direct knowledge or understand- effects of bias will have greater impact. AI to Catch Cheaters
and the results were striking: Where of deceased people. The dataset over
ing, by algorithms. It is used to create The Apple card gave higher credit limits the standard system saw “a black and AI is being used to catch cheaters. ECRI indexes on people of color, which means
advertising, to help potential employers to men than women, in some cases by a white photo of a baseball glove,” Nor- Institute’s CrossCheq uses machine that if law enforcement uses the data to
predict our behaviors, to determine our factor of 20. Wearables such as Google’s man saw “a man murdered by machine learning and data analytics to look for train algorithms, it’s going to lead to bias.
mortgage rates, and even to help law en- Fitbit are considerably less accurate for gun in broad daylight.” The point of the hyperbole and misleading information Image recognition is a particularly vexing

Society Trends
challenge, because researchers need large to turn satellite images into usable maps, headlines. Most people know about the idential candidate Andrew Yang have thiness and validity of scientific research,
datasets to perform their work. Often, was withholding certain data. Research- Uber self-driving car that hit and killed a all published works outlining different particularly in areas such as organ dona-
images are used without consent. The ers were using a neural network called pedestrian in Tempe, Arizona. But there versions of a “digital dividend”—a way tions and medical research. In addition,
Child Exploitation Image Analytics pro- CycleGAN, which learns how to map were countless more incidents that didn’t for companies to pay back to society a employing ethicists to work directly with
gram—a dataset used for testing by facial image transformations. For example, result in death, and as a result, aren’t portion of the profits derived from AI. managers and developers and ensuring
recognition technology developers—has it took an old aerial photograph of a known to the public. At the moment, developers themselves are diverse—rep-
been running since 2016 with images of neighborhood, distinguished between researchers are not obligated to report Prioritizing Trust resenting different races, ethnicities and
“children who range in age from infant streets, alleys, driveways, buildings, and accidents or incidents involving our data, genders—will reduce inherent bias in AI
through adolescent” and the majority of lampposts, and then generated a map that or AI processes, unless a law is broken. We will soon reach a point when we will systems.
which “feature coercion, abuse, and sex- could be used by GPS. Initially, they used While big companies must inform con- no longer be able to tell if a dataset has
ual activity,” according to the program’s an aerial photograph that hadn’t been sumers if their personal data—credit card been tampered with, either intentionally
own developer documentation. These seen by the network. The resulting image numbers, home addresses, passwords— or accidentally. AI systems rely on our
images are considered particularly chal- looked very close to the original—suspi- have been stolen, they are not required trust. If we no longer trust their out-
lenging for the software because of the ciously close. But on deeper inspection, to publicly document instances in which comes, decades of research and techno-
greater variability of position, context, the researchers found that many de- algorithms have learned to discriminate logical advancement will be for naught.
and more. tails in both the original image and the against someone on the basis of race or Leaders in every sector—government,
generated image weren’t visible in the gender, for example. business, nonprofits and so on—must
map made by the AI. It turns out that the have confidence in the data and algo-
AI Intentionally Hiding Data rithms used. Building trust and account-
system learned to hide information about Digital Dividends
Computers do exactly what they are told the original image inside of the image it ability requires transparency. This is a
to do. Command a machine to win at generated. Artificial intelligence will inevitably lead complicated process, and corporations,
a game, and it will do everything in its to a shift in the global workforce, causing government offices, law enforcement
power to achieve that goal. Apparently job losses across many industries. Re- agencies, and other organizations under-
that now includes cheating. Researchers Undocumented AI Accidents searchers at Oxford University’s Institute standably want to keep data private. The
at Stanford University and Google dis- Only a few of the numerous AI-relat- for Humanity, researchers at the Future ethics of how data is collected in the first
covered that an AI, which was designed ed accidents in 2018 and 2019 made Today Institute, and former U.S. pres- place may also influence the trustwor-

Bully Bots
Mid-future pessimistic scenario

We’ve seen affable chatbots inadvertently turn malicious
and inflammatory when they’re exposed to unfiltered
online conversations, but now they’re purposely pro-
grammed to antagonize. Black hat hackers cultivate bully
bots—malicious chatbots that target and troll children and
teens. Bully bots are designed to detect and exploit chil-
dren’s insecurities to inflict harm and suffering, and they
can be bought anonymously at low cost. With so much of
younger generations’ conversations taking place via texts
and DMs, there is ample data to feed the chatbots so that
they can easily impersonate their target’s peers. Ill-inten-
tioned kids—and, disappointingly, adults—use these bots
to outsource cyberbullying and deflect guilt and responsi-
bility from themselves. Opportunist criminals use them to
threaten children of prominent figures and wealthy indi-
viduals to gain access to security information and valuable
objects. With inadequate laws and frustrating layers of
anonymity, parents are at a loss over how to prevent
harm and seek justice for their victimized children.

Application
STRATEGY INNOVATION R&D RISK
Artificial intelligence affects every AI is additive to the innovation and For AI companies, talent retention is Risks abound in AI. New regulations
business across multiple dimensions. creative processes. Innovation teams likely to get worse in 2021, as com- could curb research, innovation, and
AI is a cornerstone of most organi- can use deep learning for new product panies offer increasingly competitive product development. Geopolitical
zations, from workforce automation ideation, for understanding markets, packages to attract top minds. China tensions and AI nationalism will start
to digitization to staff allocation and and for anticipating what’s on the has emerged as an R&D powerhouse, to direct foreign investment in new
beyond. It is imperative that executives horizon. Especially as no-code and posing additional talent challenges. But ways. Bias in facial recognition should
and senior managers understand what low-code applications become more for other companies, this is a good time concern everyone; so should the
AI is, what it is not, and what strategic widely available, innovation teams will to make use of commercially available quickening advancements of general
value it adds to the business. Chief build powerful systems for decision and open source frameworks, algo- adversarial networks as they are used
strategy officers should build a robust management, general brainstorming, rithms, cloud-based systems, and ap- to manipulate and generate content.
understanding of AI in order to de- and powerhouse ways of generating plications to mine your own data. There Everyone working in risk must develop
velop longer-term plans and engage new ideas. is tremendous value in learning from a sharp focus on AI. Risk models should
more closely with others in the C-suite, your customers, partners, and vendors, be developed to determine plausible
especially chief technology officers, and industry. near-future scenarios, so that leaders
chief information security officers, can adjust their strategies accordingly.
chief financial officers and others in
the organization.

Key Questions We recommend using this report to support your strategic

foresight activity in the coming year. Every executive team
should begin by asking these questions about technology,
science, and policy:
1 2 3
How well do we Are we adequately What parts of our
?
understand AI? planning for the long business model make
term? us vulnerable to
What opportunities are we
disruption brought by
leaving on the table? What assumptions must
the evolution of AI?
hold true for our current
strategy to succeed? What can we do now
to mitigate future risk?
How will we make
needed changes?

Selected Sources
A Constructive Prediction of the Gen- Representations Learned With Un- of the 2018 National Defense Strat- tent/371/6526/284
eralization Error Across Scales, ICLR supervised Pre-Training Contain Hu- egy (Department of Defense, 2018),
2020. Jonathan S. Rosenfel, Amir Ros- man-like Biases.” ArXiv abs/2010.15052 3, https://www.hsdl.org/?view&- https://blog.einstein.ai/progen/
enfeld, Yonatan Belinkov N. Shavit. (2020) did=807329
https://www.nature.com/articles/
Sharir, Or et al. “The Cost of Training Hafner, Danijar et al. “Dream to Con- President’s Council of Advisors on s41591-020-0916-2
NLP Models: A Concise Overview.” trol: Learning Behaviors by Latent Science and Technology, Recommen-
ArXiv abs/2004.08900 (2020) Imagination.” ArXiv abs/1912.01603 dations for Strengthening American https://paperswithcode.com/trends
(2020) Leadership in Industries of the Future
Artificial Intelligence Market Size, (Washington, DC: Office of Science
http://horace.io/pytorch-vs-tensor-
Share & Trends Analysis Report By Solu- “Business R&D and Innovation Survey and Technology Policy, June 2020)
flow/
tion (Hardware, Software, Services), By (BRDIS),” National Science Foundation,
Technology (Deep Learning, Machine Dario Gil, “How to Ensure the U.S.’s
https://thegradient.pub/p/
Learning), By End Use, By Region, And https://www.nsf.gov/statistics/srvyin- Quantum Future,” Scientific American,
cef6dd26-f952-4265-a2bc-f8bfb9e-
Segment Forecasts, 2020 – 2027. dustry/#sd&tabs-1. August 20, 2020, https://www.scien-
b1efb/
Grand View Research, July 2020. tificamerican.com/article/how-to-en-
Intel AI Labs: https://newsroom.intel. sure-the-uss-quantumfuture/.
https://medium.com/runacapital/
The State of Financial Crime 2021: com/press-kits/intel-labs/; NVIDIA AI open-source-growth-benchmarks-
Technology, regulation, and the future Labs: https://www.reuters.com/article/
and-the-20-fastest-growing-oss-start-
of financial crime compliance. Com- us-usa-economy-artificialintelligence/
ups-d3556a669fe6
plyAdvantage, January 2021. https://www.nvidia.com/en-in/deep- as-companies-embrace-ai-its-a-job-
learning-ai/ai-labs-her/; Uber AI Labs: seekers-market-idUSKCN1MP10D
Destructive Cyber Operations and
Machine Learning, CSET Issue Brief. https://www.uber.com/newsroom/ https://www.nature.com/articles/
Dakota Cary, Daniel Cebul. ailabs/. s41586-019-1799-6
Steed, Ryan and A. Çalişkan. “Image Department of Defense, Summary https://science.sciencemag.org/con-

Authors
LEAD AUTHOR FTI CONTRIBUTOR
Amy Webb Marc Palatucci

Founder and CEO Senior Foresight Associate
Future Today Institute Future Today Institute
awebb@futuretodayinstitute.com mpalatucci@futuretodayinstitute.com
Amy Webb pioneered FTI’s data-driven foresight University’s Säid School of Business, was elected a Marc Palatucci is a Senior Foresight Associate at the BA in Linguistics and Languages from NYU’s Gallatin
methodology that is used within hundreds of organi- life member to the Council on Foreign Relations and Future Today Institute, with research specializations School of Individualized Study. Palatucci serves as
zations globally. Her focus is to transform organiza- is a member of the Bretton Woods Committee. She in new realities (AR/ VR/ MR/ DR), digital fashion, editor-at-large for an arts, fashion and culture mag-
tions as they prepare for complex futures. She advises is a member of the World Economic Forum where future of work and learning, retail and e-commerce. azine and collaborates with writers, designers and
CEOs of the world’s most-admired companies, three- she serves on the Global Future Council on Media, He co-leads the MBA course in strategic foresight at producers on films, TV shows and commercials about
star admirals and generals, and the senior leadership Entertainment and Culture and the Stewardship the New York University Stern School of Business. science, technology and the future.
of central banks and intergovernmental organiza- Board of the Forum’s Platform for Shaping the Fu- Palatucci serves on the World Health Organization’s
tions. She leads FTI’s technology research initiatives ture of Media, Entertainment and Culture. Webb is Learning Strategy Advisory Group and is a Senior
on AI, synthetic biology and genetic engineering, the bestselling author of several books about strategic Deputy to the World Economic Forum’s Platform for
next-generation networks, and quantum technol- foresight and emerging technologies. A lifelong sci- Shaping the Future of Media, Entertainment and Cul-
ogies. Webb is a professor of strategic foresight at ence fiction fan, she collaborates closely with writers ture. He holds an MBA in Emerging Technology from
the New York University Stern School of Business, and producers on films, TV shows and commercials New York University’s Stern School of Business and a
where she developed and teaches the MBA course on about science, technology and the future.
strategic foresight. She is a Visiting Fellow at Oxford

Authors
FTI CONTRIBUTORS EDITORIAL
Leah Zaidi Jennifer Alsever Cheryl Cooney Tom Foster

Senior Foresight Associate Editorial Director Director of Operations Editor
Future Today Institute Future Today Institute Future Today Institute
lzaidi@futuretodayinstitute.com jalsever@futuretodayinstitute.com ccooney@futuretodayinstitute.com Sarah Johnson
Editor
Leah Zaidi is an award-winning futurist with special- Jennifer Alsever is the Future Today Institute’s For over a decade, Cheryl Cooney has served as the
izations in systems thinking, worldbuilding, science Editorial Director. She has been a journalist for more Future Today Institute’s director of operations, where
fiction prototyping and experiential futures. She is an than two decades covering tech, biotech, startups and she manages workflows, planning and logistics.
Associate Editor of the World Futures Review. She business for such publications as Fortune Magazine, Cooney is a published poet, with works appearing in
holds an MDES in Strategic Foresight and Innova- the Wall Street Journal, the New York Times, Wired American and Australian anthologies.
tion from OCAD University and a BA from York and Fast Company. A popular young adult fiction
University. writer, Alsever has won multiple YA awards for her
Trinity Forest Series.

Authors
CREATIVE BUSINESS DEVELOPMENT
Emily Caufield Julia Durgee Mel Blake

Creative Director Artist and Futurist Business Development
Future Today Institute mblake@futuretodayinstitute.com
ecaufield@futuretodayinstitute.com
Emily Caufield is an award-winning designer and Julia Durgee is a rare left and right-brained MBA with Mel Blake handles commercial development and
illustrator. Serving as FTI’s creative for more than a creative, strategic, analytical, and digital experience at client relations for FTI. He was formerly founder
decade, Caufield applies design thinking to visually world-class brands. She illustrated the portraits that and managing director of Monitor Talent, a speaker
communicate complex trends, scenarios and foresight appear in this year’s trend report. agency he founded at the Monitor Group, a global
research. Caufield designed all aspects of this year’s consulting firm. He is a board member of The Yale
trend report. She is a graduate of Boston University’s Center for Customer Insights. He holds an MBA
College of Fine Arts. from the Yale School of Management and a BA in
Economics from Wesleyan University.

Why FTI
We answer your most challenging questions using data, We support executive leaders and their teams.
creative inquiry, and strategic foresight.
The Future Today Institute works closely with executive leadership and management
teams to transform their strategic thinking on the future. Advisory services include
• What are plausible deep (20+ years), long-range (10+ years), and near-term
signal mapping, trend identification, scenario development, risk modeling, visioning,
(2+ years) futures?
and strategic planning.
• What scenarios describe our futures?
• What’s happening outside my industry that I should know?
• What companies, startups, and partners make up our future value network?
• What new products, services, or businesses should we build?
• Which tech trends should we monitor? When should we act?
• How can we build an early warning system to see the next disruptive event?
• How do we reduce uncertainty about our futures?

About the Future Today Institute
Founded in 2006, the Future Today Institute researches, mod- Contact Us

els, and prototypes future risk and opportunity. As the lead- The Future Today Institute
ing strategic foresight and futures management consultants 33 Irving Place
to executive leadership teams worldwide, FTI’s data-driven 10th Floor
New York, NY 10003
applied research reveals trends and calculates how they will
disrupt business, government, and society. hello@futuretodayinstitute.com
267-342-4300
Together with our clients and partners, FTI is helping leaders
www.futuretodayinstitute.com
achieve their preferred futures. Our pioneering, data-driven
forecasting methodology and tools empower leaders to make
better decisions about the future, today.

Disclaimer
The views expressed herein are the authors’ own and are not representative of the
greater organizations in which they have been employed. The names of companies,
services, and products mentioned in this report are not necessarily intended as en-
dorsements by the Future Today Institute or this report’s authors.
The Future Today Institute’s 2021 Tech Trends Report relies on data, analysis, and
modeling from a number of sources, which includes sources within public and private
companies, securities filings, patents, academic research, government agencies,
market research firms, conference presentations and papers, and news media stories.
Additionally, this report draws from the Future Today Institute’s previous EMT Trends
Reports, FTI Trend Reports, and newsletters. FTI’s reports are occasionally updated on
the FTI website.
FTI advises hundreds of companies and organizations, some of which are referenced
in this report. FTI does not own any equity position in any of the entities listed in this
presentation.
Any trademarks or service marks used in this report are the marks of their respective
owners, who do not endorse the statements in this report. All rights in marks are re-
served by their respective owners. We disclaim any and all warranties, expressed or
implied, with respect to this report.

Using and Sharing The Material In This Report

You are free to: Under the following terms: You are prohibited from:
Share Attribution Commercial Sharing

Copy and redistribute the material You must give appropriate credit Don’t copy and redistribute this
in any medium or format, including to the Future Today Institute, pro- material in any medium or for-
in your organizations and class- vide a link to this Creative Com- mat for commercial purposes,
We invite you to use, share, and build
rooms. mons license, and indicate if any including any personal/ corporate
upon the material in our 14th annual
changes were made. You may do marketing and client services.
Future Today Institute Tech Trends Re-
Adapt so in any reasonable manner, but
port. We are making it freely available
Remix, transform, and build upon not in any way that suggests that Representing This Work As
to the public. This work is licensed under
the material for your own research, the Future Today Institute endors- Your Own
a Creative Commons Attribution-Non-
work, and teaching. es you or your use. Do not represent any part of this
Commercial-ShareAlike 4.0 Interna-
tional License. material as your own without
NonCommercial giving credit to the Future Today
You may not, under any circum- Institute.
stance, use the material for com-
mercial purposes. Additional Restrictions
You may not apply legal terms
ShareAlike or technological measures that
If you remix, transform, or build legally restrict others from doing
upon the material, you must dis- anything this license permits.
tribute your contributions under
the same license as you see here.

00 01 02 03 04 05 06 07 08 09 10 11 12
This is volume 1 in the Future Today Institute’s

2021 Tech Trends Report. Each volume covers
a different set of topics.
To find additional volumes, visit

www.futuretodayinstitute.com/trends
SEE MORE REPORT HERE:

VIETNAM MARKET
REPORT COMMUNITY

Volume 1-Tech - Trends - Volume - 1 - AI - Shared - by - VN - Market - Report - Community

Uploaded by

Copyright:

Available Formats

You might also like

Volume 1-Tech - Trends - Volume - 1 - AI - Shared - by - VN - Market - Report - Community

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Volume 1-Tech - Trends - Volume - 1 - AI - Shared - by - VN - Market - Report - Community

Uploaded by

Copyright:

Available Formats

Volume 1 of 12

14th Annual Edition

03 Overview 19 Algorithm Marketplaces 32 Machine Image Completion 43 Nation-based Guardrails

03 © 2021 Future Today Institute

Macro Forces and

Our framework organizes nearly 500 trends into 12 clear categories.

Each trend offers six important insights.

Macro Forces and

Scenarios Describe Plausible Outcomes

2. Temporal and Emotive Tags 3

05 © 2021 Future Today Institute

+ Natural language processing is + Natural language processing

+ The race is on to capture AI + New artificial nervous systems

+ Gaussian processes, the gold + Technical alliances that help

+ Several countries will launch

+ New measures to regulate the

07 © 2021 Future Today Institute

14TH YEAR ON THE LIST

08 © 2021 Future Today Institute

09 © 2021 Future Today Institute

10 © 2021 Future Today Institute

12 © 2021 Future Today Institute

13 © 2021 Future Today Institute

Emotion AI Will words we say: We express ourselves

Still Needs to Work interact with those around us.

AI may be the answer to preserv-

© 2021 Future Today Institute

15 © 2021 Future Today Institute

16 © 2021 Future Today Institute

17 © 2021 Future Today Institute

18 © 2021 Future Today Institute

19 © 2021 Future Today Institute

Rage Against the Machine

Mid-future neutral scenario

20 © 2021 Future Today Institute

Health, Medicine, and Science Trends

22 © 2021 Future Today Institute

Health, Medicine, and Science Trends

23 © 2021 Future Today Institute

Health, Medicine, and Science Trends

24 © 2021 Future Today Institute

Deep Twins in the OR

Mid-future optimistic scenario

25 © 2021 Future Today Institute

27 © 2021 Future Today Institute

card companies, banks, local authorities, Personal Digital Twins

28 © 2021 Future Today Institute

Code is important for reproducibility, Facebook seems to have gained ground.

of $1 per 1000 parameters. OpenAI’s IBM

Papers published at NeurlIPS 2019 (fractional count)

31 © 2021 Future Today Institute

32 © 2021 Future Today Institute

33 © 2021 Future Today Institute

34 © 2021 Future Today Institute

AI Brain Drain like New York University professor Yann

versities. In some cases, professors take uate-level, research-based AI university.

36 © 2021 Future Today Institute

37 © 2021 Future Today Institute